Journal of Networks - CiteSeerX

Journal of Networks ISSN 1796-2056

Volume 5, Number 11, November 2010 Contents Special Issue: All-Optically Routed Networks

Guest Editors: Antonio Teixeira, Anna Tzanakaki, Davide Careglio, and Miroslaw Klinkowski

Guest Editorial Antonio Teixeira, Anna Tzanakaki, Davide Careglio, and Miroslaw Klinkowski

SPECIAL ISSUE PAPERS Trigonometric Transforms for High-Speed Optical Networks: All-Optical Architectures and Optical OFDM Michela Svaluto Moreolo, Valentina Sacchieri, Gabriella Cincotti, and Gabriel Junyent Analyzing Power Consumption in Optical Cross-connect Equipment for Future Large-Capacity Optical Networks Makoto Murakami A Simple Generalized Approach to Node Failure Recovery with Span-Protecting p-Cycles Diane Prisca Onguetou and Wayne D. Grover Impairment Aware RWA in Optical Networks: Over-provisioning or Cross Optimization? Konstantinos Christodoulopoulos, Panagiotis Kokkinos, Konstantinos Manousakis, and Emmanouel A. Varvarigos ICBR-Diff: an Impairment Constraint Based Routing Strategy with Quality of Signal Differentiation Amornrat Jirattigalachote, Paolo Monti, Lena Wosinska, Kostas Katrinis, and Anna Tzanakaki Feedback Based Load Balancing, Deflection Routing and Admission Control in OBS Networks Sébastien Rumley, Christian Gaumier, Oscar Pedrola, and Josep Solé Pareta Design and Development of a Semantic Information Modelling Framework for a Service Oriented Optical Internet Chinwe E. Abosi, Reza Nejabati, and Dimitra Simeonidou Reducing Complexity and Consumption in Future Networks G. M. Tosi Beleffi, G. Incerti, L. Porcari, S. Di Bartolo, M. Guglielmucci, A. L. J. Teixeira, L. Costa, N. Wada, J. Prat, J. Lazaro, and P. Chanclou Decrease of the Link PMD by Fiber Exchange and Investigation of the PMD Distribution along Buried Optical Fibers with a POTDR Armin Ehrhardt, Manuel Paul, Lars Schürer, Christoph Gerlach, Wolfgang Krönert, Daniel Fritzsche, Dirk Breuer, Volker Fürst, Normand Cyr, Hongxin Chen, Gregory W. Schinn

Deployment and Validation of GMPLS-Controlled Multi-layer Integrated Routing over the ASON/GMPLS CARISMA Test-bed Fernando Agraz, Luis Velasco, Jordi Perelló, Marc Ruiz, Salvatore Spadaro, Gabriel Junyent, and Jaume Comellas Enhancing Performance of Optical Communication Systems with Advanced Optical Signal Processing Ivan Glesk, Marc Sorel, Anthony E. Kelly, and Paul R. Prucnal Advanced Test-beds to Validate Physical Estimators in Heterogeneous Long Haul Transparent Optical Networks Annalisa Morea, Florence Leplingard, Jean-Christophe Antona, Pascal Henri, Thierry Zami, and Daniel C. Kilper All-optical Label Swapping Techniques for Optical Packets at Bit-rate Beyond 160 Gb/s Nicola Calabretta, Hyun-Do Jung, and Harm Dorren Tb/s Transmission and Routing Systems Using Integrated Micro-Photonic Components Efstratios Kehayas, Leontios Stampoulidis, and Paraskevas Bakopoulos

REGULAR PAPERS A Peer-to-Peer Game Model using Punishment Strategies Chunzhi Wang, Hongwei Chen, Ke Zhou, Hui Xu, and Zhiwei Ye Reliable Resource Search in Scale Free Peer-to-Peer Network Wei Song, Wenbin Hu, Zhengbing Hu, and Xi Zeng Tree Routing Protocol with Location-based Uniformly Clustering Strategy in WSNs Gengsheng Zheng and Zhengbing Hu A SVM Method for P2P Traffic Identification based on Multiple Traffic Mode Hongwei Chen, Xin Zhou, Fangping You, Hui Xu, Chunzhi Wang, and Zhiwei Ye

Special Issue on All‐Optically Routed Networks

Guest Editorial The evolution of communication networks is driven by a huge increase of end user traffic and bandwidth demands due to

the recent massive deployment of broadband access technologies as well as the outburst of new network applications and emerging service oriented applications. The observed trends are accompanied by the advances in optical technologies which have enabled the development of long‐distance and high‐capacity transmission systems. Nowadays the role of optics in communication networks is mainly restricted to transmission and some limited switching functions. However, next generation optical networks are foreseen to perform extended switching and control processing operations in the optical domain such as fast provisioning of high bandwidth services, dynamic management of end‐to‐end resources in an efficient and reliable manner, support of a wide range of service granularities, design and development of power‐efficient protocols and equipment, guarantee of both quality of service and transmission for various user applications, and support of network and user data security. Apart from answers to existing technology challenges, the development of all‐optically routed networks imposes the need for novel approaches and solutions to meet the requirements dictated by the evolving and new services that are becoming available to the end users. In this context, several issues need to be addressed so that to overcome the limited scalability and flexibility of today’s network infrastructures as well as insufficient network manageability and increased overall capital and operational expenditure leading to high network services costs.

Born as a small conference in Poland, the International Conference on Transparent Optical Networks (ICTON) is today a reference meeting point thanks to the initiative and continuous support of Prof. Marian Marciniak. ICTON, now at its 11th edition, gathers each year around 500 researchers from 5 continents: experts in different fields like optical processing, optical networking, nanophotonics, photonic crystals, photonic components, etc. interchange their knowhow and experience with novel researchers and PhD students. This special issue presents fourteenth selected papers from ICTON 2009 addressing most of such heterogeneous topics. The papers are organized in three groups according to the approach followed i.e. physical layer modeling and simulations, networking aspects (design, provisioning and protocols) and experimental demonstrations and test-beds.

The first group consists of two papers. The paper entitled Trigonometric transforms for high-speed optical networks: all-optical architectures and optical OFDM

by M. Svaluto et al. discuss the use of all-optical discrete Hartley transform (DHT) and discrete Cosine transform (DCT) architectures for high-speed optical signal processing, filtering in optical communication systems which in some cases can advantageously replace the Fourier transform for both all-optical and electronic signal processing. The authors instantiate the benefits of the proposed methodology through an orthogonal frequency division multiplexing (OFDM) system example. Furthermore, the implications and inherent benefits of the approach are further developed.

The paper entitled Analyzing power consumption in optical cross-connect equipment for future large-capacity optical networks by M. Murakami focuses on energy consumption of the equipment used in WDM optical networks. The equipment under consideration and comparison are optical cross-connects supporting Tb/s capacities based either on electronic or photonic switching. Results indicate that the electronic solution introduces power consumption requirements exceeding 50kW, while the photonic counterpart requires 8-30kW. In the photonic switching case the power requirements are mainly determined by the power consumption of the associated transponders.

The second group consists of five papers. The paper entitled A simple generalized approach to node failure recovery with span-protecting p-cycles by D. P.

Onguetou and W. D. Grover proposes a new strategy based on the concept of p-cycles to recover a single node failure in optical networks. This new strategy, called two-hop flow, is based on a generalization of how nodes in an ordinary p-cycle derive survivability through loop-back at the nearest two neighbor-nodes on the same cycle. An ILP formulation as well as a faster heuristic algorithm for the two-hop strategy is given. The performance of the two-hop flow method is finally compared with other p-cycle-based node failure recovery methods such as Node-Encircling p-Cycles, Flow-Protecting p-Cycles, and Failure-Independent Path-Protecting p-Cycles.

The second and third papers of this group address a similar problem, i.e. how to incorporate the physical layer impairments constraints into the network protocol decisions. The paper entitled Impairment aware RWA in optical networks: over-provisioning or cross optimization? by K. Christodoulopoulos et al. focuses on the evaluation of the overall network performance following two different impairment aware routing approaches: one considering worst case physical performance assumptions independent of the actual status of the network at the time of routing and a second considering the current network utilization and perform a cross layer optimization between the network and physical layers. Simulation results indicate that the second approach provides improved network performance at the expense of increased complexity, which however does not significantly increase the corresponding execution times. The paper entitled ICBR-Diff: an Impairment Constraint Based Routing Strategy with Quality of Signal Differentiation by A. Jirattigalachote et al. propose a

JOURNAL OF NETWORKS, VOL. 5, NO. 11, NOVEMBER 2010 1245

novel impairment constraint based routing (ICBR) algorithm with differentiation of services based on the BER of a lightpath. Presented results reveal that significant network performance improvement in terms of connection blocking can be achieved, compared to non-differentiated conventional RWA and ICBR algorithms.

The paper entitled Feedback based load balancing, deflection routing and admission control in OBS networks by S. Rumley et al. focuses on Optical Burst Switching (OBS) and proposes a unified scheme supporting adaptive deflection routing and admission control to handle various types of traffic variations aiming at simplifying the OBS network architecture and enhancing its flexibility. Simulation results evaluate the performance of the proposed solution and compare it to traditional deflection routing indicating specific benefits.

The last paper of this group entitled Design and development of a semantic information modelling framework for a service oriented optical Internet by C. E. Abosi et al. focuses on the new concept of service plane architecture which is considered as an architectural enhancement promising to simplify the management of heterogeneous, dynamic and complex emerging service requirements in distributed IT systems. In particular, the authors propose a rich descriptive semantic framework and common vocabulary for the description of network and IT resources.

The last group consists of seven papers. The paper entitled Reducing complexity and consumption in future networks by G. Tosi-Beleffi et al. reports the main

results of the EU FP7 SARDANA project which targets the performance enhancement of dense Fibre-to-the-Home networks by providing large bandwidth to the end user in a flexible and intelligent way. Main features of the SARDANA network proposal are described, namely remote amplification for network reach extension, remote monitoring of the network infrastructure, and remote ONUs powering.

The paper entitled Decrease of the link PMD by fiber exchange and investigation of the PMD distribution along buried optical fibers with a POTDR by A. Ehrhardt et al. reports field trial results of the cumulative PMD distribution in deployed fiber sections which are obtained using a novel POTDR prototype instrument. The presented measurement technique can be applied both for old fibers in order to replace bad pieces selectively and for newly installed cables directly after installation to check the quality and the adherence of the fibers to predetermined physical limits. This information is particularly important for network operators who want to improve their networks in order to install systems at 40 Gbit/s and beyond.

The paper entitled Deployment and validation of GMPLS-controlled multi-layer integrated routing over the ASON/GMPLS CARISMA test-bed by F. Agraz et al. addresses the problem of grooming connection requests with fine granularity into optical channels. Firstly, the authors design a GMPLS-controlled multi-layer architecture with grooming-capable transport network. Then, they test experimentally the proposed multi-layer solution highlighting its good trade-off between network blocking probability and E/O port usage when compared to all-optical and opaque solutions.

The paper entitled Enhancing performance of optical communication systems with advanced optical signal processing by I. Glesk et al. describes two applications of advanced optical signal processing techniques. On one hand, optical XOR gates are used to apply one-time pad encrypted algorithm ensuring data security directly in the optical layer. On the other hand, an ultra fast optical signal conditioning technique is introduced in an OCDMA system to improve its scalability. Both signal processing techniques are successfully demonstrated in experimental platforms. Currently, the authors are working on the monolithic integration of such techniques.

The paper entitled Advanced test-beds to validate physical estimators in heterogeneous long haul transparent optical networks by A. Morea et al. present novel diverse experimental set-ups required in transparent optical networks to replace the old point-to-point test-bed set-ups. In particular, the authors propose a Quality of Transmission (QoT) estimator to evaluate the accumulation of the physical layer impairments in a lightpath and verify its feasibility by emulating the networks with double loop experiments. The proposed QoT estimator is assessed and refined through the realization of diverse experiments under different network uncertainties (eg. Dispersion maps, fiber types heterogeneity and number of neighbor channels travelling).

The paper entitled All-optical label swapping techniques for optical packets at bit-rate beyond 160 Gb/s by N. Calabretta et al. present two different techniques based on optical signal processing to realize a scalable all-optical packet switch with label swapping. For both techniques, the authors report experimental results showing the routing operation of the 160 Gb/s packets and beyond. A comparison between the techniques in terms of devices, bit-rate scalability, latency, power consumption, power penalty performance and cascadability is also made.

The last paper entitled Tb/s transmission and routing systems using integrated micro-photonic components by E. Kehayas et al. report recent advances in the development of photonic switching and transmission systems that exploit high and low index contrast integration materials. Based on functional examples of the refereed techniques, the authors group the applications in two sets: the transmission/regeneration prone and the more compact and efficient which are more prone to photonic routing platforms. The diverse presented technologies are enablers for the future systems-on-chip.

The Guest Editors would like to thank the authors for their high-quality contributions. Special thanks go to Prof. Marian

Marciniak, the Conference Organizer of ICTON 2009, for facilitating and supporting the preparation of this special issue.

1246 JOURNAL OF NETWORKS, VOL. 5, NO. 11, NOVEMBER 2010

Guest Editors: Antonio Teixeira, Institute of Telecommunications/DETI, Universidade de Aveiro, Portugal Anna Tzanakaki, Athens Institute of Technology, Greece Davide Careglio, Universitat Politècnica de Catalunya, Spain Mirosław Klinkowski, National Institute of Telecommunications, Poland

Antonio Teixeira (http://www.it.pt/person_detail_p.asp?ID=703) is an associate professor at the Electronics, telecommunications and informatics department of the Universidade de Aveiro, researcher at the Insitituto the Telecomunicações and Senior specialist at the industrial environment team for standardization in Nokia Siemens Networks. He received the BsC and MsC in electronics engineering and telecommunications, in July 1994 at U. Aveiro and the Ph.D. degree in electrical engineering in 1999 also from the University of Aveiro but partly developed at the institute of Optics, U. of Rochester, NY, USA. He was an engineer at the research branch of Portugal Telecom in 95 and visitor researcher at several institutes like NICT in Japan. He is a co-author of over 300 publications in international journals and conferences and 6 chapters in books. He is a co-inventor of 7 national and international patents. He is a member of the IEEE and participated in several Technical Program Committees and has chaired some national and international conferences. His research interests include optical access networks, all-optical processing, radio over fiber, network consolidation, monitoring. He has a long list of participations in national and international projects

where he acted both as participant or leader.

Anna Tzanakaki is an Associate Professor at the Athens Information Technology, where she is leading the Network Design and Services research group. She is also an adjunct faculty member of Carnegie Mellon University, USA. She has obtained a BSc degree from the University of Crete, Greece, an MSc and a PhD both from the University of Essex, UK. She was a co-founder and a senior engineer of ilotron ltd, a spin-off from the University of Essex, involved in the design of systems for WDM optical networks. Following ilotron, she joined Altamar Networks, a subsidiary of Ditech Communications, as a principal engineer responsible for optical architecture and system design. She is a co-author of over 100 publications in international journals and conferences. She is a co-inventor of 1 granted and 11 published patents. She is a senior member of the IEEE and several Technical Program Committees. Her research interests include optical wavelength, burst and packet switched networks, cross-layer network design and traffic provisioning

as well as network convergence in support of telecommunications and IT services.

Davide Careglio (http://people.ac.upc.edu/careglio) is an Associate Professor in the Department of Computer Architecture at the UPC, Barcelona, Spain. He received the M.Sc. (2000) and Ph.D. (2005) degrees in Telecommunications Engineering both from UPC. He also received the Laurea degree in Electrical Engineering from Politecnico di Torino (2001). Since 2000 he is a staff member of the CCABA (Advanced Broadband Communication Center). His research interests are in the fields of networking protocols with emphasis on optical switching technologies, and algorithms and protocols for traffic engineering and QoS provisioning. He is a co-author of over 80 publications in international journals and conferences. He has participated in many European and national projects in the field of optical networking and green communication. He is a member of the IEEE and participated in the Technical Program Committees of several conferences, including IEEE ICC and IEEE Globecom.

Mirosław Klinkowski is an Assistant Professor in the Department of Transmission and Optical Technology at the National Institute of Telecommunications in Warsaw, Poland, and is a Collaborating Researcher at the Universitat Politècnica de Catalunya (UPC), Barcelona, Spain. He received his M.Sc. and Ph.D. degrees, respectively, from Warsaw University of Technology in 1999 and from UPC in 2008. His publications include several book chapters and more than 50 papers in relevant research journals and refereed international conferences. He has participated in many European projects dealing with topics in the area of Optical Networking. He is currently involved in the COST 2100 action. His research interests include optical and wireless networking with emphasis on network modeling, design, and performance analysis.

Trigonometric Transforms for High-SpeedOptical Networks: All-Optical Architectures and

Optical OFDM

Michela Svaluto MoreoloCentre Tecnologic de Telecomunicacions de Catalunya (CTTC), Castelldefels (Barcelona), Spain

Email: michela.svaluto@cttc.es

Valentina Sacchieri, Gabriella CincottiDepartment of Applied Electronics, University Roma Tre, Rome, Italy

Email: {vsacchieri, cincotti}@uniroma3.it

Gabriel JunyentDepartment of Signal Theory and Communications, Universitat Politecnica de Catalunya (UPC), Barcelona, Spain.

Centre Tecnologic de Telecomunicacions de Catalunya (CTTC), Castelldefels (Barcelona), SpainEmail: gabriel.junyent@cttc.es

Abstract— In this paper, it is shown that the use of realtrigonometric transforms can advantageously replace theFourier transform in optical communication systems, forboth all-optical and electronic signal processing. All-opticaldiscrete Hartley transform (DHT) and discrete Cosinetransform (DCT) architectures for high-speed optical signalprocessing, filtering and data compression are given andcompared to fiber Fourier processing. A trigonometrictransform-based orthogonal frequency division multiplexing(OFDM) transmission system is also presented. The modu-lation/demodulation is performed by using the fast Hartleytransform (FHT) algorithm. The simplified scheme and thereal processing make it suitable for direct detection opticalsystems.

Index Terms— discrete Hartley transform, discrete Fouriertransform, discrete cosine transform, optical signal process-ing, orthogonal frequency division multiplexing

I. I NTRODUCTION

The deployment of high-speed, large-capacity opticalsystems, able to support the high growth of IP traffic isaddressing new technical challenges. As the transmissionspeed increases, signal degradation issues and transmis-sion impairments, such as chromatic dispersion (CD) andPolarization Mode Dispersion (PMD), severely limit theattainable distance. On the other hand, the electronicbottleneck represents an impairment to achieve higherthroughput and fully exploit the optical bandwidth.

Suitable signal processing plays a fundamental rolein designing efficient and cost-effective solutions for

This paper is based on “Signal Processing Based on TrigonometricTransforms for High-speed Optical Networks,” by M. Svaluto Moreolo,V. Sacchieri, and G. Cincotti, which appeared in the Proceedings of the11th IEEE International Conference on Transparent Optical Networks(ICTON), Island of Sao Miguel, Azores, Portugal, June 28-July 2 2009.c© 2009 IEEE.

flexible, high-capacity optical networks. Processing sig-nals directly in the optical domain provides bit-rateand signal-format independent transmission schemes, pre-serving end-to-end optical transparency; while a signalprocessing in the electrical domain takes advantage ofthe mature technology and capabilities of digital signalprocessing (DSP).

In this paper, we present transform-based signalprocessing for high-speed optical networks, showing thatreal trigonometric transforms, e.g. discrete Hartley trans-form (DHT) and discrete Cosine transform (DCT), canoutperform standard discrete Fourier transform (DFT) forboth all-optical and electronic signal processing.

DHT and DCT furnish a recursive approach to easilydesign higher order optical transforms. Compared to DFToptical architectures, they can be implemented by simplerall-optical passive circuits, that can be used for theanalysis, compression and filtering of optical signals.

Furthermore, we present an Optical Orthogonal Fre-quency Division Multiplexing (O-OFDM) transmissionsystem based on Hartley transform. The processing per-formed in the electrical domain takes advantage of thetrigonometric transform algorithm, without dealing withthe complex processing of the standard DFT-based imple-mentation. The OFDM scheme can be simplified and itresults suitable for direct detection optical systems.

The paper is organized as follows: in Section II, wepresent the all-optical trigonometric transforms and theirproperties for signal processing in the optical domain.In Section III, we describe the modulation/demodulationscheme based on Hartley transform for optical OFDMsystems. We analyze the spectral behaviour and the com-putational complexity. Finally, conclusions are given inSection IV.

x(N/2-1)

.N/2-order FHT

N/2-order FHT ...

ho(0)x(N/2)

x(N-1)

ho(N/2-1)

he(N/2-1)

Figure 1. Block diagram of the recursive procedure to opticallyimplement the N-order FHT of a sequencex(n); at the output the evenand odd components of the transformed sequence are in bit reverse order,he(k) and ho(k) respectively.

II. TRIGONOMETRIC TRANSFORMS FORALL-OPTICAL SIGNAL PROCESSING

Hartley transform and cosine transform are real trigono-metric transforms and can be advantageously used inoptical communications exploiting their properties foroptical signal processing. They are powerful tools forsignal filtering and data compression. The possibility toevaluate both transforms directly in the optical domainis particularly attractive for high-speed optical signalprocessing.

The N -order DHT of a sequencex(n) is defined asfollows

h(k) =1√N

N−1∑n=0

x(n)[cos(2πkn/N) + sin(2πkn/N)]

k = 0, 1, · · · , N − 1. (1)

It is suitable for the analysis of real signals, since theHartley transform of a real signal is real, while stan-dard Fourier transform performs a complex processingand the phase always carries fundamental information.Indeed, the DFT kernel,exp(−j2πkn/N), can be writtenas cos(2πkn/N) − j sin(2πkn/N); it differs from theDHT kernel,cos(2πkn/N) + sin(2πkn/N), only in theimaginary unit. Accordingly, the realR(k) and imaginaryX(k) parts of the Discrete Fourier Transform (DFT),f(k), coincide with the even and the negative odd parts ofthe Discrete Hartley Transform (DHT),E(k) and−O(k),respectively:

f(k) = R(k) + jX(k) = E(k)− jO(k). (2)

SinceE(k) = [h(k) + h(N − k)]/2 andO(k) = [h(k)−h(N − k)]/2, the DHT is suitable to calculate powerspectra. The power spectrum ofx(n) can be evaluatedasS2(k) = R2(k) + X2(k), by using the DFT, or moreeasily and efficiently asS2(k) = [h2(k)+h2(N − k)]/2,by using the DHT.

Moreover, many theorems of the Fourier transformcan be applied to the Hartley transform, in some casesafter slight modifications [1]. The convolution theorem

Figure 2. All-optical circuit for the 8-order FHT. The three blocks inthe figure indicate the 3dB couplers bank, the 4-order FHT circuit andthe additional odd block stage merged with the 4-order FHT, accordingto the scheme in Fig. 1.

for DHT has key relevance in image processing anddigital filtering. In fact, when one of the two signals tobe convolved presents even symmetry, the DHT of theconvolution is simply the product between the DHT ofthe two signals themselves. Therefore, DHT can be ad-vantageously used in place of DFT, to avoid calculationsbetween complex values. Likewise the cross-correlationtheorem for DHT does not deal with complex conjugationand the product is performed between real values [1].

Complex computations are carried out in optics by theuse of phase shifters that are not required in the opticalimplementation of DHT. In terms of matrix formulation,the FFT differs from the FHT only in a matrix term, whichis diagonal with imaginary values (−j) in the Fourierprocessing and is a unitary permutation matrix in the caseof real DHT processing [2]. The optical circuit imple-menting an N-order FHT can be simply synthesized withan optical passive fiber network of asymmetric couplers,without the need of phase shifters, as required for theoptical implementation of FFT. As shown in the blockdiagram of Fig. 1, the N-order FHT optical circuit canbe easily derived with a recursive procedure [3], startingfrom the simplest circuit synthesizing a DHT of secondorder [

h(0)h(1)

[1 11 −1

] [x(0)x(1)

]. (3)

This matrix formulation evidences that a simple 3dBcoupler, with scattering matrix

M50/50 =1√2

[1 11 −1

], (4)

gives the DHT of the input values. Arranging the output inbit reverse order, the matrix formulation of greater ordertransform is derived from the half-length order trans-form. It only requires an additional stage of asymmetriccouplers for the odd block. The matrix values furnishthe corresponding asymmetric couplers for the opticalimplementation and the optical equivalent architecture canbe simply synthesized [3].

As an example, we analyze the case of the FHT oforder N = 8. The 3dB couplers bank stage consistsof four elements synthesized by four asymmetric50/50beam splitters, described by the scattering matrix (4). Bysplitting the outputs values according to their parity, thematrix formulation can be written as follows

h(0)h(4)h(2)h(6)

=H4√

x(0) + x(4)x(2) + x(6)x(1) + x(5)x(3) + x(7)

andh(1)h(5)h(3)h(7)

=H4√

1 0 0 00 1 0 00 0 α α0 0 α −α

x(0)− x(4)x(2)− x(6)x(1)− x(5)x(3)− x(7)

where α = 1/√

2 and H4 is the normalized matrixdescribing the transform of orderN/2 = 4

1 1 1 11 −1 1 −11 1 −1 −11 −1 −1 1

Both the input and output vectors in (5) and (6) areordered according to their parity and each array of inputvalues is given by the half-length vector at the output ofthe 3dB couplers bank. The transform circuit is depictedin Fig. 2. The odd stage is implemented by a single3dB coupler and two fibers; it has been merged withthe following optical circuit of the4-order FHT. Thesame stage of the optical FFT circuit requires three phaseshifters (see Fig. 3 of Ref. [4]).

Similarly, we can derive the procedure to opticallyimplement theN -order DCT of a sequencex(n), whichis defined as

z(k) =

b(k)N−1∑n=0

x(n) cos[π(2n + 1)k/2N ]

k = 0, 1, · · · , N − 1, (8)

b(k) ={ 1√

2k = 0

1 0 < k ≤ N − 1.(9)

The DCT can be evaluated through fast algorithms byusing direct or indirect computations. The former ap-proach needs algorithms with less computation steps, buta rigorous procedure is not easy to provide or to beproved; the latter is based on fast algorithms of othertransforms (e.g. FFT, FHT) and takes advantage of theirefficiency [6], [7]. DCT is suitable for data compression,since it concentrates the signal energy only in its lowindex coefficients. This is due to the energy compactionproperty, that allows neglecting the high indices withoutaffecting the content of the signal. Moreover, the DCT isalso useful in optical image reconstruction, filtering andfeature extraction [8].

The DCT can be advantageously calculated from theDHT, with a faster and simpler implementation compared

x(N/2-1)

N-order FHTodd block

N/2-order DCT

zo(0)x(N/2)

x(N-1)

tterfl

zo(N/2-1)

ze(N/2-1)

Figure 3. Block diagram of the recursive procedure to opticallyimplement the N-order DCT of a sequencex(n).

50/5050/50

4/9650/50

50/50�

70/30�

� 50/50�

Figure 4. All-optical circuit to evaluate the 8-order DCT. The threeblocks in the figure indicate the 3dB couplers bank, the 4-order DCTcircuit and the 8-order FHT odd stage merged with the DCT additionalbutterfly stage, according to the scheme in Fig. 3.

to indirect computation from FFT: complex processingcan be avoided and no phase shifters are required for theoptical implementation. The N-order DCT of a sequencex(n) can be simply derived by the N-order FHT of thesequencex(n), with x(n) = x(2n) and x(N − n− 1) =x(2n + 1) [5]. It follows that the design of the N-orderDCT optical circuit, by means of the DHT-based indirectcomputation, is very simple. The first step consists ininverting the input elements, in order to use the FHToptical architecture of the same order. Then the butterflystage has to be designed and added at the output of theFHT odd block, according to the matrix formulation of[5]. It is also possible to derive a recursive procedure,described in Fig. 3: after combining the reordered inputswith a 3dB couplers bank, the even block is processedby the DCT circuit of orderN/2 and the odd block isevaluated by an optical circuit obtained from the modifiedFHT network. In Fig. 4, we give an example for the DCTcircuit of order N = 8. The DCT architecture can becompared with the circuit of Fig. 2: both the even andodd blocks of the two optical transforms differ for theadditional butterfly stages. The couplers elements havebeen derived from the matrix formulation values. The

Figure 5. DFT frequency responses for different values of k; the spectraamplitude has been normalized.

butterfly stage of the4-order DCT circuit has been mergedto the odd block element of the4-order DHT, giving asingle coupler with scattering matrix

M15/85 =[

sin(π/8) cos(π/8)cos(π/8) − sin(π/8)

]. (10)

The optical DCT circuit has been further optimized byinverting the couplers position in the odd block [2]. Thebutterfly stage is given by two asymmetrical couplers withscattering matrixM4/96 andM70/30.

III. H ARTLEY TRANSFORM FOR OPTICALOFDM

OFDM is a multi-carrier transmission technique, wherethe signal data stream is transmitted over several lower-rate sub-channels, whose sub-carriers are orthogonal toeach other, so that their spectra partially overlap. The useof OFDM in optical networks mitigates the transmissionimpairments and also enables to transmit high data rate.The high tolerance to CD and PMD allows extending theoptical transmission reach to thousands of kilometers [9],[10].

The signal processing in the OFDM transmitter/receivertakes advantage of the efficient algorithm of FFT, whichenables the use of available DSP devices. The OFDMsignal is then modulated on an optical carrier; directdetection and coherent schemes can be used, tradingsimplicity against increased sensitivity. These alternativesolutions result from the critical issue of transmitting thecomplex bipolar OFDM signal in an optical system.

We present an alternative OFDM scheme, for op-tical networks based on the FHT, which can advan-tageously replace the FFT to implement the OFDMmodulation/demodulation, as demonstrated for high-speedwireless communications [11].

A. Spectral analysis

In order to study the DHT-based OFDM modulation,we analyze the frequency response of the DFT and theDHT of a unitary symbol sequence overN = 32 points.The spectra are shown in Fig. 5 and 6 for different OFDM

Figure 6. DHT frequency responses for different values of k; the spectraamplitude has been normalized.

sub-carriers, by varying the value of the parameterk.A band of 5GHz for the OFDM sub-channels has beenconsidered. Due to the DHT kernel, two mirror-symmetricsub-bands carry each symbol of the data sequence, as it ispossible to appreciate from Fig. 6, especially for the sub-carriers corresponding tok = 10 andk = 20. Frequencysub-channels separation and orthogonality are preserved,and since the DHT spectra are split into mirror-symmetricsub-bands, the frequency diversity is enhanced.

B. Optical DHT-based OFDM scheme

We exploit the properties of the real trigonometrictransform to streamline the conventional OFDM scheme,based on FFT, and achieve a simplified transmissionsystem. If the input data sequence is mapped into a realconstellation, the inverse FHT (IFHT) gives real values,so that only the in-phase component (I), and no imaginarycontribution (Q), has to be processed, as shown in Fig. 7.

At the receiver side, the parallel processing is per-formed by using a forward FHT and the received con-stellation symbols are demapped. The direct and inverseHartley transforms are identical, so that the same DSPdevice can be used at the transmitter and at the receiver.

Two digital-to-analog (DACs) and and two analog-to-digital (ADCs) converters are required when OFDMmodulation/demodulation is based on standard FFT: onefor the real and one for the imaginary part of the complexOFDM signal. When the OFDM processing is based onDHT, only one single DAC and one single ADC arerequired, as in discrete multi-tone modulation (DMT)systems [12]. The DMT is a multi-carrier transmissiontechnique considered a special case of OFDM, where noI/Q modulation at radio frequency (RF) is needed andthe number of required electronic devices is reduced. InDMT systems, the OFDM signals are real-valued: theinput sequence is forced to have Hermitian symmetryso that its inverse FFT gives real values. Thanks to theDHT real processing, the Hermitian symmetry constrain isnot required in the proposed OFDM system. Real OFDMsignals are simply obtained if the transmitted bit sequenceis mapped into a real constellation.

DHT-OFDM TX

DHT-OFDM TXDHT-OFDM TX

OFDM RX

OFDM RXOFDM RX

inputdata

output data

FilterModulator Fiber link

Optical domain

Electricalamplifier

Figure 7. Optical DHT-based OFDM system. The OFDM-modulation and demodulation are performed in the electrical domain by using an N-orderinverse FHT and forward FHT, respectively. The optical detection is direct. The components in grey are in the optical domain.

As depicted in Fig. 7, the optical system is very simple:one single Mach-Zehnder modulator is required for trans-mitting the signal that can be easily recovered by usingdirect detection (DD). A simple photodetector is used,which can be modeled with a square law characteristic.The photocurrent is given by the OFDM signal termand unwanted mixing products that can be minimized byinserting a guard band between the OFDM signal and theoptical carrier [13]. This is at expenses of the spectralefficiency [14]. Single-side band (SSB) modulation canbe adopted to ensure that the OFDM subcarriers arerepresented only once by the optical frequencies [15]. Ifthis solution is considered, an optical filter is requiredto transmit in the fiber channel only one side of theoptical spectrum, that is symmetric with respect to theoptical carrier. In Fig. 7 the additional component isindicated with dashed line. To reduce the ASE noiseanother optical filter with narrow band can be addedbefore the photoreceiver [15].

C. Computational Complexity

The computational complexity of FHT is lower thanthe FFT of complex data. Only real multiplications haveto be calculated and no complex algebra has to be applied[5], [16]. Therefore, FHT is faster than FFT, based on theCooley and Tukey algorithm [17].

Compared to optimized algorithms to evaluate the FFTof real-valued sequences, the FHT algorithms requireabout the same number of multiplications but more ad-ditions [18]. For example, in the case of decimation-in-time or decimation-in-frequency radix-2 algorithm,N−2more additions are required to evaluate anN -order DHT,compared to the real-valued DFT of the same order.

Similarly for radix-4, split radix, prime factor andWinograd transform algorithms, the number of additionsrequired by the FHT slightly exceeds the ones requiredby the FFT of a real-valued sequence [18]. The algorithmwith minimum arithmetic complexity implementing theDHT requires only two more additions than the FFT al-gorithms for real-valued signal with the minimum number

of multiplications and additions [19]. When DSP devicesare used, this improvement increases the computationalspeed.

Moreover, to obtain a real-valued sequence, the inverseFFT requires a complex vector with Hermitian symmetry.This results in additional computational resources, com-pared to the FHT algorithms.

Finally, the same FHT routine can be applied to cal-culate the DHT and the IDHT, because they are equal,as mentioned in Sec. III-B. This is an advantage overany algorithm to evaluate the DFT, which requires anadditional control to reverse the sign in the transformkernel.

IV. CONCLUSION

In this paper, it has been shown that all-optical passivearchitectures implementing DHT and DCT are powerfultools for high-speed optical signal processing, filteringand data compression. Compared to DFT, which dealswith a complex processing, DHT performs a simplifiedfiltering of real signals and some properties can be advan-tageously applied in image processing and digital filtering.DHT optical implementation requires a fiber optical net-work of asymmetric couplers without phase shifters andwe have shown that the indirect computation of opticalDCT based on DHT circuits furnish a simpler design anda recursive approach to easily implement higher ordertransforms. Furthermore, we have presented an opticalOFDM transmission system based on Hartley transform.We have analyzed the frequency response in comparisonto the DFT-based approach. A simplified scheme has beengiven, resulting suitable for optical DD systems. If theinput sequence is mapped with a real constellation, theDHT-OFDM signal is real. Only one single DAC/ADCis used for the modulation/demodulation, as in DMTsystems, and the Hermitian symmetry constrain is notrequired. The computational complexity has been alsoanalyzed in comparison to DFT algorithms for complexand real-valued data.

ACKNOWLEDGMENT

This work was founded by the MICINN (SpanishMinistry of Science and Innovation) through the projectDORADO (TEC2009-07995) and developed within theBONE-project (“Building the Future Optical Networkin Europe”), a Network of Excellence funded by theEuropean Commission through the 7th ICT-FrameworkProgramme.

REFERENCES

[1] R. N. Bracewell, “Discrete Hartley transform,” J. Opt. Soc.Amer., vol. 73, pp. 1832-1835, Dec. 1983.

[2] M. Svaluto Moreolo, G. Cincotti, “Fiber Optics Trans-forms,” Proc. Int. Conf. Transparent Optical Networks, IC-TON 2008, 22-26 June 2008, Athens, Greece.

[3] M. Svaluto Moreolo, G. Cincotti, “Fiber Hartley Transformand Optical Indirect Computation of Discrete Cosine Trans-form,” IEEE Trans. Comm., to appear.

[4] A. E. Siegman, “Fiber Fourier optics,” Opt. Lett., vol. 19,pp. 12151217, 2001.

[5] H. S. Hou, “The fast Hartley transform algorithm,” IEEETrans. Computers, vol. C-36, pp. 147-156, Feb. 1987.

[6] P.Lee and F. Huang, “Restructured recursive DCT and DSTalgorithms,” IEEE Trans. Signal Processing, vol. 42, pp.1600-1609, 1994.

[7] H. S. Malvar, “Fast computation of the discrete cosinetransform through the fast Hartley transform,” Electron. Lett.,vol. 22, pp, 352-353, 1986.

[8] A. V. Oppenheim, R. W. Schafer, with J. R. Buck,Discrete-time signal processing, Prentice Hall, Upper Saddle River,New Jersey 1999.

[9] A. J. Lowery, and J. Armstrong, “Orthogonal-frequency-division multiplexing for dispersion compensation of longhaul optical systems,” Optics Express, vol. 14, pp. 2079-2084,2006.

[10] S. L. Jansen, I. Morita, T. C. W. Schenk, N. Takeda, andH. Tanaka, “Coherent optical 25.8-Gb/s OFDM transmissionover 4160-km SSMF,” J. Lightwave Technol., vol. 26, pp.6-15, Jan. 2008.

[11] D. Wang, D. Liu, F. Liu and G. Yue “A novel DHT-basedUltra-Wideband System,” Proc. ISCIT 2005, vol. 50, pp. 172-184, June 2004.

[12] S. C. J. Lee, F. Breyer, S. Randel, H. P. A. van denBoom, A. M. J. Koonen, “High-speed transmission overmultimode fiber using discrete multitone modulation,” J. Opt.Networking, vol. 7, pp. 183-196, Feb. 2008.

[13] W. Shieh, I. Djordjevic, OFDM for optical communica-tions, Elsevier, USA, 2010.

[14] J. Armstrong, “OFDM for optical communications,” J.Lightwave Technol., vol. 27, pp. 189-204, Feb. 2009.

[15] B. J. C. Schmidt, A. J. Lowery and J. Armstrong, “Ex-perimental demonstration of electronic dispersion compensa-tion for long-haul transmission using direct-detection opticalOFDM,” J. Lightwave Technol., vol. 26, pp. 196-203, Jan.2008.

[16] R. N. Bracewell, “The fast Hartley transform,” Proc. IEEE,vol. 72, pp. 1010- 1018, Aug. 1984.

[17] J. W. Cooley, and O. W. Tukey, “An Algorithm for theMachine Calculation of Complex Fourier Series,” Math.Comput., vol. 19, pp. 297-301, 1965

[18] H. V. Sorensen, D. L. Jones, C. S. Burrus, M. T. Heideman,“On Computing the Discrete Hartley Transform,”IEEE Trans.Acoust., Speech, Signal Processing, vol. ASSP-33, pp. 1231-1238, Oct. 1985.

[19] P. Duhamel, M. Vetterli, “Improved Fourier and Hartleytransform algorithms: application to cyclic convolution of realdata,” IEEE Trans. Acoust., Speech, Signal Processing, vol.ASSP-35, pp. 818-824, June 1987.

Michela Svaluto Moreolo received the M.Sc. degree withhonours in Electronics Engineering and the Ph.D. degree inTelecommunications Engineering from University Roma Tre,Rome, Italy, in May 2003 and April 2007, respectively. Dur-ing her PhD, she has been a visiting researcher at Instituteof Semiconductor and Solid State Physics, Johannes KeplerUniversity, Linz, Austria. From January 2007 to December 2008she has held a Postdoctoral position at the Applied ElectronicsDepartment of University Roma Tre. Currently she is withthe Optical Networking Area of CTTC, as a Research Asso-ciate. Her research interest areas are optical signal processingand advanced transmission technologies for high-speed opticalnetworks, including optical OFDM, multiplexing and codingtechniques.

Valentina Sacchieri received the M.S. degree in ElectronicsEngineering from University of Roma TRE, Rome, Italy, inOctober 2006. She is currently a Ph.D. student in Telecommuni-cation Engineering at the University of Roma TRE, Rome, Italy,since November 2006. She was hosted as visiting researcherat the Telecommunication Institute (IT) of Aveiro University,Portugal, from January to June, 2008. Her research interestsare in the field of next generation optical access networks andare focused on OCDMA techniques, optical scrambling andOCDMA networks confidentiality, optical OFDM and advancedmodulation formats. She is a student member of the IEEE, IEEEPhotonics Society, IEEE Communications Society, GTTI and amember of CNIT.

Gabriella Cincotti received the Laurea (M. Sc.) degree cumlaude in Electronic Engineering from “La Sapienza” Universityof Rome, in April 1992. She was a Project Engineer at theMicrowave laboratory of ALENIA, Aeritalia & Selenia S.p.A.,in Rome, from 1992 to 1994. She joined the Department ofElectronic Engineering of University “Roma Tre”, as AssistantProfessor, in October 1994. In May 2005 she became AssociateProfessor at the Department of Applied Electronics. Her researchinterests are in optical packet switching and optical code divisionmultiple access networks and devices. Prof. Gabriella Cincottihas been the Guest Editor of IEEE/OSA Journal of LightwaveTechnology /Optical Signal Processing 2006, and serves asTopical Editor of OSA Optics Letters. Prof. Gabriella Cincottihas authored over 200 papers in refereed journals and conferenceproceedings, holds a Japanese and two international patents.She is a Senior Member of IEEE Lasers and Electro-OpticsSociety (LEOS), and member of the National Inter-UniversityConsortium for Telecommunications (CNIT), the Optical So-ciety of America (OSA) and Inter-University Consortium forMatter Science (CNISM).

Gabriel Junyent is a telecommunication engineer (UniversidadPolitecnica de Madrid, UPM, 1973), and holds a Ph.D. degreein communications (Universitat Politecnica de Catalunya, UPC,1979). From 1973 to 1989, he was a teaching assistant andassociate professor at the UPC, where he has been a fullprofessor since 1989. In the last 15 years he has participatedin more than 30 national and international R&D projects, andhas published more than 30 journal papers and book chaptersand 100 conference papers.

WDMline

clientingress

clientegress

Fig. 1 Optical network consisting of optical cross-

connects with 4 express and 1 add/drop branches in which OEO converters are used depending on route.

Analyzing Power Consumption in Optical Cross-connect Equipment for Future Large-Capacity

Optical Networks

Makoto Murakami NTT Network Service Systems Laboratories, 180-8585 Tokyo, Japan

E-mail: murakami.makoto@lab.ntt.co.jp

Abstract—We describe a comparative analysis of the power consumption of optical cross-connect (OXC) equipment based on electrical and/or photonic matrix switching, which will be used for future large-capacity optical networks. The switch configurations used for both types of OXCs are also comprehensively discussed. For optical networks that accommodate traffic with a capacity of several Tb/s, the power consumption of the OXC equipment based on electrical switching could reach more than 50 kW; this can be reduced to 8 – 30 kW if photonic switching is used instead. In photonic switching-based OXC equipment, the power consumption of transponders becomes the most significant, and its reduction is the key issue for power-efficient large-capacity OXC equipment. Index Terms— transparent optical network, power consumption, optical cross-connect, optical switch

I. INTRODUCTION

In general, as the traffic in a network increases, the power/energy consumed to transmit a broadband signal becomes greater. This means that we will suffer from the effects of significant power/energy consumption because of the rapid broadband-traffic growth that is expected to occur worldwide. A large increase in power consumption in telecommunication networks may lead to stringent requirements for saving energy to reduce climate change, i.e., the current accelerated warming of the planet thought to be due to the release of man-made greenhouse gases, as is true in other industries.

Information communications technology (ICT) industries are far from innocent in the matter of climate change. However, the ICT sector actually does not contribute a large amount of greenhouse gas emissions compared to its share of the global gross domestic product (GDP) because the primary sources of greenhouse gases are energy production and consumption, transport, buildings, and so on. ICT may also have a positive impact on climate change through use of computing and telecommunications networks, e.g., reducing carbon emissions of other sectors and offering a climate-change monitoring system. Therefore, most standardization organizations including ITU-T have intensively discussed the power consumption/saving issues of optical transmission equipment [1].

In addition to the environmental problem, a significant increase in power consumption could have a large impact on the operational expense of network operators. Therefore, specifications and proper management of power consumption in network systems are critical for coping with this situation. In addition, some telecommunication carriers have started making their own standards to measure the power consumption of

telecommunications equipment [2], and accordingly, metrics and test procedures to determine the power consumption of core routers have also been developed [3].

Next-generation optical networks are expected to offer a large traffic capacity to any destination at any time. Such networks will be achieved by using ring and/or mesh architecture and optical cross-connects (OXCs) linked to each other by wavelength-division multiplexing (WDM) transmission lines, as shown in Fig. 1. The OXC equipment should have a large number of input-output ports, e.g., more than 100, and handle wavelength signals of 10 Gb/s or larger to accommodate the expected significantly growing number of broadband users in the future. One type of OXC is based on electrical switching and is widely used as an established technology. The second type using photonic switching is expected to enable large-capacity networks without a large increase in power consumption.

We analyze and compare the power consumption of future large-capacity OXCs, one based on electrical

TABLE I. SWITCH CONFIGURATION

Configuration Non-blocking type Number of switch elements

Crossbar Wide sense N2

Clos Strict sense 4Sqrt(2)N1.5-4N

Benes Re-arrangeable N(2log2N-1)/2

Spanke Strict sense 2N(N-1)

switching and the other based on photonic switching. First, we will discuss configurations of electrical and photonic switches for large-capacity OXCs. Then, we calculate the power consumption of both types of OXCs and discuss the results with a view to further reduce power consumption for future large-capacity networks.

II. SWITCH CONFIGURATION

A. Electrical-switching-based OXC The configurations of typical OXCs are illustrated in

Fig. 2. The OXCs should have express ports, which flexibly connect a wavelength signal in an ingress WDM line to any egress WDM line. In addition, the OXCs have client ports, which drop/add any wavelength signal to/from an express port for transmission in a WDM line. Figure 2 shows an OXC based on electrical matrix switching technology, which consists of optical amplifiers, optical multiplexers (MUX)/demultiplexers (DEMUX), and optical-electrical (O/E) and electrical-optical (E/O) converters for express and client signal accommodation.

A large-sized electrical matrix switch is usually achieved by integrating small switch elements. Typical configurations of integrated switch elements are shown in Table 1, and their arrangements are shown in Fig. 3 [4]. The number of switch elements required for configuring each type of switch is shown in Fig. 4.

The most basic configuration is the crossbar configuration, a simple integration of 2 × 2 switch elements as shown in Fig. 3 (a). This configuration has an

advantage in that no blocking occurs in a wide sense but the increase of the number of switch elements required is expressed by the square of the number of switching ports, as shown in Fig. 4. The second one is the Clos configuration, which uses k switches having m ports in its input and output sides to configure an N × N large switch, where N = k × m. This configuration also has non-blocking of ports and fewer switching elements than the crossbar one; the increase of switching elements in this configuration is expressed by N to the power of 1.5. The Benes configuration can drastically decrease the number of switch elements because the increase is expressed by N × (2log2N-1)/2 as shown in Fig. 4. However, this type of configuration requires some re-arrangement of the switching connection when blocking occurs. The Spanke

configuration is configured with 1 × N switches and has an advantage of non-blocking characteristics. However, if the 1 × N switches are created by 2 × 2 switches, the number of switch elements is as large as that of the crossbar configuration; its increase is estimated as the order of N square.

Surveying actually available commercial OXC equipment, it seems the Benes configuration is the most useful and commonly adopted in many places. This

electrical matrixswitch

client add

client drop

To WDM line

From WDM line

opticalamplifier

opticalMUX

opticalDEMUX

O/EO/EO/EO/E

E/OE/OE/OE/O

O/E E/O

Fig. 2 Optical cross-connect configurations based on electrical switching.

0 200 400 600 800 1000

Number of ports

CrossbarClosBenes

Fig. 4 Number of switch elements required to create each type of configuration.

p x m (a) (b)

N x 1 (C) (d)

Fig. 3 Typical switch configurations: (a) crossbar, (b) Clos, (c) Benes, (d) Spanke.

TABLE II. TECHNOLOGIES FOR PHOTONIC SWITCH

Technology

Dimensions Wave guide Free space

2 Mach-Zehnder

interferometers on PLC

MEMS mirrors for on-off switching

3 Not available MEMS mirrors with angle control in 3-D

arrangement

means that the power consumption of a switching fabric itself used in the OXC equipment may increase by the order of N × log2N. However, an actual switch fabric may have other components that do not rapidly increase with the switching elements, and the increase of power consumption can be estimated as larger than linear but less than N × log2N.

B. Photonic-switching-based OXC The use of photonic switching in OXCs is expected to

contribute to reducing power consumption in future large-capacity networks because photonic switches do not depend on signal speed, while electrical switches require more power for higher speed signals. Moreover, it may lead to reduction in power consumption through the elimination of transponders if most wavelength signals are transparently cross-connected. However, it seems difficult to construct an optical network with only photonic switching, i.e., fully transparent. Figure 5 shows an OXC based on photonic matrix switching technology, which consists of optical amplifiers, optical MUX/DEMUX, and optical-electrical-optical (OEO) converters for client signal accommodation and wavelength conversion and/or extension of reach. The number of OEO converters for wavelength conversion will depend on the difficulty in transparently passing through the wavelength signals caused by wavelength resource conflicts in the WDM line at the egress ports.

The technologies to create a photonic switch are listed in Table 2. For 2-dimensional switches, the most promising way is to use planar lightwave circuit (PLC) technology. A typical PLC-based photonic switch can be composed of Mach-Zehnder interferometers acting as a 2 × 2 switch and uses the crossbar configuration. The power consumption of each switch element depends on the current changing the optical path delay of one of the arms of the interferometer to switch the optical output port. Typical power consumption was reported as 0.15 W per Mach-Zehnder interferometer [5]. In actual

fabrication, the PLC switch elements are two Mach-Zehnder interferometers cascaded to attain sufficiently low crosstalk, and the power consumption of each switch element is double that of the Mach-Zehnder interferometer. It has been recently reported that a lower power consumption of 20 mW per interferometer was achieved in laboratory experiments [6].

Another type of 2-dimensional photonic switch uses free space optics with small mirrors fabricated by micro-electro-mechanical systems (MEMS) technology and has the benefit of free space optics, including lower crosstalk and typical power consumption as small as 5 mW.

When those 2-dimensional switch elements are integrated in the crossbar configuration, the total power consumption of an N × N crossbar switch reaches N2/2 times that of each switch element, assuming that half of them are in the on-state. Moreover, it is still too difficult to create a larger size switch, e.g., over 100 × 100, with such a 2-dimensional configuration because of its physical size and difficulty in fabrication.

Instead, we can find that the most significant advantage of photonic technology over electronics is the availability of 3-dimensional free space optics. Thus, the most promising approach to large size photonic switches is the 3-dimensional MEMS switch shown in Fig. 6, which directly connects any input port to an output port by controlling the tilt angles of two oppositely located MEMS mirrors [7-10]. The angles of the MEMS mirrors are determined by the balance between the torsion of the spring at the mirror hinge and the electro-static force driven by the high voltage control circuit. Basically no electrical current flows to maintain the mirror angle, and the only component consuming power is the circuit for creating the control signal with a high voltage of the order of 100 V. Moreover, unlike 2-dimensional switches,

control/drive circuit

optical input

optical output

control signal

Fig. 6 Configuration of 3D MEMS switch.

photonic matrixswitch

client add

client drop

O/E/OO/E/OO/E/OO/E/O

To WDM line

From WDM line

opticalamplifier

opticalMUX

opticalDEMUX

O/E/OO/E/OO/E/OO/E/O

Wavelength converters

Fig. 5 Optical cross-connect configurations based on photonic switching.

the number of switch elements in a 3-dimensional switch linearly increases and the driving circuits may be simply added according to the port increase. The features above mean that this type of switch inherently has an advantage in reducing power consumption in large capacity OXCs.

III. ANALYSIS OF POWER CONSUMPTION

We need to create a generic OXC equipment model to

calculate the power consumption. Figure 6 shows a typical schematic rack/shelf model of OXC equipment in which each shelf can be further divided into units or cards. The main rack may contain an electrical or photonic switch unit, an optical multiplexer and demultiplexer unit, and optical amplifiers. The additional sub-racks contain transponders (O/E/O in Fig. 5) as wavelength converters or receivers (O/E in Fig. 2), and transmitters (E/O) for express/client signal accommodation. We can then assign a typical power consumption value to each card or unit by referring to current manufacturers’ actual products. For example, typical power consumption values in the O/E or O/E/O components of current OXC or WDM equipment may range from 50–100 W/card.

We calculated the power consumption of optical cross-connect equipment based on electrical switching (E-OXC) and photonic switching (P-OXC). Both types of equipment were assumed to have 10-Gb/s interface cards in the transmitters/receivers or transponders. Figure 7 shows the power consumption calculated against the number of ports, which is the sum of the express and client ports (one side).

In the case of an E-OXC, we assume the power consumption of the electrical switch increases almost linearly with the number of ports or the capacity accommodated. It is quite difficult to predict exactly the increase of the power consumption of the actual switch

fabric because of many factors other than simple integration of the switch elements. However, this assumption of linear increase will not deviate by more than a factor of log2N, which is about 6.6 and 9.9 for N = 100 and 1000, respectively, if the switch fabric uses the Benes configuration. Figure 7 shows that the power consumption increases with the number of ports and reaches over 10 kW if the number of ports is more than 160, i.e., the capacity handled in the equipment is 1.6 Tb/s and will increase to over 50 kW when the number of ports is 1000 or 10 Tb/s in capacity. Even this value, which may be conservatively estimated, could be difficult to accept for network operators because the power consumption of current large-capacity core routers, which is in the range of 10 kW, is already problematic for telecom carriers.

We then calculated the power consumption of a P-

OXC for various configurations depending on the number of wavelength converters. The power consumption of the photonic switch fabric is assumed to linearly increase according to the formula 0.48×(port counts-80)+35 [W], which is created to fit currently available large-capacity

power unit

MUX/DEMUX

common/supervisory

main rack

switch

power unit

common/supervisory

transponders (OEO)

sub rack

transponders (OEO)optical amps.

Fig. 6 Rack mount models of optical cross-connect equipment.

20,000

40,000

60,000

80,000

100,000

0 200 400 600 800 1,000

Number of ports

P-OXC w/owavelengthconvertersP-OXC with 25%wavelengthconvertersP-OXC with 50%wavelengthconvertersP-OXC with 75%wavelengthconvertersP-OXC with 100%wavelengthconverters

Fig. 7 Power consumption of various optical cross-connect configurations based on electrical and photonic switching

technologies.

3D-MEMS-based photonic switches. When wavelength routing does not cause any conflict in a WDM line and no wavelength converter is equipped in the P-OXC, the power consumption could be drastically reduced compared to the E-OXC in all calculated ranges of the number of ports. The power consumption value of such a P-OXC is only 8 kW, which is almost 1/7 of that of an E-OXC, even if the number of ports is over 1000. This means that such a P-OXC could be a 10-Tb/s-class OXC with power consumption lower than that of current 1-Tb/s-class large core routers. However, the wavelength routing that avoids any wavelength resource conflict in a WDM line without wavelength conversion is not realistic if the network scale and number of nodes are not small.

Therefore, we calculated the power consumption of a P-OXC with varying numbers of wavelength converters equipped, as shown in Fig. 7. Some increase in the number of ports is inevitable for P-OXCs with less than 100% wavelength converters, while no additional ports are necessary for P-OXCs with 100% wavelength converters. The number of ports for wavelength converters is expressed as the percentage of the number of express ports in Fig. 7. It must be noted that the number of ports described in Fig. 7 does not include the increase in the actual port count in the matrix switch

fabric for wavelength conversion, which may not be visible outside the equipment.

As clearly shown in Fig. 7, the power consumption of a P-OXC increases with the percentage of wavelength converters. However, even when 100% wavelength converters are equipped in a P-OXC, the power consumption values are still under those of an E-OXC. This is because the photonic switch consumes less power than the electrical switch. A P-OXC with a percentage of wavelength converters of around 25 – 50% could be realistic from the viewpoint of network operation. Then the power consumption of P-OXCs could be calculated in the range of 15 – 30 kW, even for a 10-Tb/s capacity.

Figure 8 shows the detailed contribution of each component to the power consumption of E-OXC and P-OXC equipment for 32 – 800 ports. For every case, the rate of contribution is almost constant if the number of ports increases to above 100. This means that the number of components is almost proportional to the number of ports. For a P-OXC, the contribution of the photonic switch is quite small compared with that of the other components, even for the one with no wavelength converters. In contrast, for an E-OXC, the electrical switch significantly contributes about 20% or more to power consumption. It is also noted that the most

32 100 200 400 600 800

Number of ports

ion Optical amp.

Express (O/Eand E/O)Client (O/Eand E/O)Electrical SW

OpticalMUX/DEMUXCommon

32 100 200 400 600 800

Number of ports

ion Optical amp.

Client (O/E/O)

Photonic SW

OpticalMUX/DEMUX

Common

(a) (b)

32 100 200 400 600 800

Number of ports

ion Optical amp.

Wavelengthconverters(O/E/O)Photonic SW

Client (O/E/O)

OpticalMUX/DEMUX

Common0%

32 100 200 400 600 800

Number of ports

Optical amp.

Wavelengthconverters(O/E/O)Photonic SW

Client (O/E/O)

OpticalMUX/DEMUX

Common

(c) (d)

Fig. 8 Element contributions to equipment power consumption in percentages: (a) electrical switching, (b) photonic switching without wavelength converters, (c) photonic switching with wavelength converters occupying 25% of

express ports, (d) photonic switching with wavelength converters occupying 50% of express ports.

significant contributions are made by the O/E, E/O, and OEO devices for express and client signal accommodation or for wavelength conversion. In particular, for a P-OXC without wavelength converters, the O/E/O devices for client signal accommodation significantly contribute more than 70%. The impact of the O/E/O devices is also significant for a P-OXC with wavelength converters, as shown in Figs. 8(c) and (d). Those results indicate that the key to reducing the power consumption of P-OXCs is to improve the power consumption efficiency of the O/E/O components.

The advantage of P-OXCs over E-OXCs, i.e., the optical transparency of the switch fabric, becomes salient when the optical signal speed increases to 40- or 100-Gb/s. The power consumption of the photonic switch fabric may not change even for such high-speed signals and the switching capacity can be simply increased with the optical signal speed, while that of the electrical switch fabric incurs significant increase in power consumption as the total switching capacity increases with the signal speed. In terms of the transponder or O/E/O converter, the power consumption has not necessarily increased in proportion to the signal speed so far, and thus it seems that the use of higher speed signals leads to reduction in power consumption per bit. However, modulation and demodulation schemes for higher speed signal transmission of 40 or 100 Gb/s use optical phase and/or polarization multiplexing instead of simple intensity modulation commonly used for 10-Gb/s systems [11]. Such a complicated scheme may cause inevitable increase in the number of discrete optical and electrical components in the O/E/O converters, and reduction in power consumption per bit may be difficult to achieve. The key technology to solve this problem will be the development of a power-efficient photonic integration circuit [12].

IV. CONCLUSION

Optical cross-connect (OXC) equipment is essential for future optical networks, which are expected to offer large-capacity traffic to any destination at any time. The use of photonic switching in OXC equipment instead of conventional electrical switching is effective in reducing power consumption, and this will be emphasized as the optical signal speed increases and larger switching capacity is accommodated. These advantages of photonic switching may result from free-space technology in 3-dimensional configurations. However, it seems that placing some O/E/O converters in the equipment is necessary in actual complicated optical mesh networks in order to settle wavelength resource conflict in a WDM transmission line and to meet the requirement for smaller switching granularity than the wavelength signal. The power consumption of such a realistic OXC configuration then significantly depends on the O/E/O conversion devices, more than 70% depending on the rate of inclusion. We expect further reduction in power consumption per bit to be obtained by increasing the optical signal speed to 40 or 100 Gb/s and developing

higher-speed transmission technology. In such a high-speed signal region, complicated modulation/demodulation schemes will be inevitable and may not necessarily contribute to reducing power consumption effectively, i.e., reduction to less than 4 or 10 times the power of current 10-Gb/s transponders. Thus, lower-power-consuming transponders (O/E/O devices) obtained by developing un-cooled devices and photonic integration are expected to further reduce the power consumption in OXCs for future large-capacity optical networks.

REFERENCES

[1] http://www.itu.int/ibs/ITU-T/200809climate/ [2] http://www.verizonnebs.com/TPRs/VZ-TPR-9205.pdf [3] D. Kharitonov, B. Nordman, and A. Alimian, “Network

and Telecom Equipment — Energy and Performance Assessment,” Focus group ICT&CC, September 2008.

[4] R. Ramaswami and K. N. Sivarajan, “Optical Networks,” Academic Press, 2002.

[5] T. Goh, A. Himeno, M. Okuno, H. Takahashi, and K. Hattori, “High-extinction Ratio and Low-loss Silica-based 8×8 Strictly Nonblocking Thermo-optic Matrix Switch,” IEEE J. Lightwave Technol., Vol. 17, No. 7, pp. 1192–1199, 1999.

[6] K. Watanabe, Y. Hashizume, Y. Nasu, M. Kohtoku, M. Itoh, and Y. Inoue, “Ultralow Power Consumption Silica-Based PLC-VOA/Switches,” IEEE J. Lightwave Technol., Vol. 26, No. 14, pp. 2235–2244, 2008.

[7] J. Kim, C. J. Nuzman, B. Kumar, D. F. Lieuwen, J. S. Kraus, A. Weiss, C. P. Lichtenwalner, A. R. Papazian, R. E. Frahm, N. R. Basavanhally, D. A. Ramsey, V. A. Aksyuk, F. Pardo, M. E. Simon, V. Lifton, H. B. Chan, M. Haueis, A. Gasparyan, H. R. Shea, S. Arney, C. A. Bolle, P. R. Kolodner, R. Ryf, D. T. Neilson, and J. V. Gates, “1100×1100 Port MEMS-based Optical Crossconnect with 4-dB Maximum Loss,” IEEE Photon. Technol. Lett., Vol. 15, No. 11, pp. 1537–1539, Nov. 2003.

[8] X. Zheng, V. Kaman, S. Yuan, Y. Xu, O. Jerphagnon, A. Keating, R. C. Anderson, H. N. Poulsen, B. Liu, J. R. Sechrist, C. Pusarla, R. Helkey, D. J. Blumenthal, and J. E. Bowers, “Three-dimensional MEMS Photonic Cross-connect Switch Design and Performance,” IEEE J. Sel. Top. Quantum Electron., Vol. 9, No. 2, pp. 571–578, 2003.

[9] J. Yamaguchi, T. Yamamoto, N. Takeuchi, and A. Shimizu, “Free-space Optical Interconnection System with MEMS Mirrors,” in EOS Topical Meeting Optics in Computing, 2004, pp. 96–97.

[10] M. Murakami, T. Seki, and K. Oda, “Optical Signal Channel Power Stability in Transparent Optical Network using Large-scale Photonic Crossconnects and Automatic Gain Control EDFAs,” J. Opt. Commun. Networks, Vol. 2, No. 1, pp. 20–27, 2010.

[11] A. Sano, E. Yamada, H. Masuda, E. Yamazaki, T. Kobayashi, E. Yoshida, Y. Miyamoto, R. Kudo, K. Ishihara, and Y. Takatori, “No-Guard-Interval Coherent Optical OFDM for 100-Gbps Long-Haul WDM Transmission,” IEEE J. Lightwave Technol., Vol. 27, No. 16, pp. 3705–3713, 2009.

[12] M. J. Wale, “Photonic Integration Challenges for Next-Generation Networks,” ECOC2009, Paper 1.7.4, Vienna, 2009.

A Simple Generalized Approach to Node Failure Recovery with Span-Protecting p-Cycles

Diane Prisca Onguetou and Wayne D. Grover TRLabs, 1200 Harley Court, 10045-111 Street, T5K 2M5, Edmonton, AB, Canada

ECE Dept., University of Alberta, 2nd Floor ECERF, 9107-116 Street, T6G 2V4, Edmonton, AB, Canada {donguetou, grover}@trlabs.ca

Abstract—This paper shows that, viewed in a generalized “two-hop” framework for node failure recovery, p-cycles actually have a very high inherent ability to restore paths transiting through a failed node. We also showed that with relatively little, if any, extra spare capacity, the principle is also amenable to explicit design of networks for 100% node and span failure protection with a single efficient set of p-cycles that support both functions. This is very different than the often-prevailing assumption that “ordinary” p-cycles offer no node protection, or only the same node-protection as a BLSR ring embodies. Indeed, the two-hop paradigm for recovery of affected paths transiting failed nodes could provide an attractive option for future network operators in that “ordinary” p-cycles are more localized, fast acting, and simple to plan and operate than any other option such as NEPCs, flow-protecting p-cycles or FIPPs. Index Terms—ordinary p-cycles, two-hop, node failure protection, R1-node.

I. INTRODUCTION Network survivability design is primarily focused on

recovery from span failures because the frequencies of fiber cable cut events are hundreds to thousands of times higher than corresponding reports of transport layer node failures. Nevertheless much less frequent than span failures, node outages are particularly harmful when they arise, at least because each specific node failure involves the simultaneous failure of all node-incident edges. On the other hand, if optical-cross-connects (OXCs) tend to be highly robust and well protected, IP/MPLS routers still suffer downtimes about as frequently as span failures because of software’ patches, upgrades or even crashes. Thus, it is not utopian to consider protection against both span and node failures in the design.

Doing so, it has been known that end-to-end path-protecting architectures such as shared backup path protection (SBPP), demand-wise shared protection (DSP) and pre-cross-connected trails (PXT) inherently provide some protection against intermediate node failures arising somewhere along the working paths. Corresponding levels of node failure restorability depend on backup channel-capacities and node-disjointness considerations in the shared risk link groups (SRLGs). In contrast to path-oriented paradigms, span-protecting architectures are based on the deployment of a set of backup path-segments between the end-nodes (i.e., the “custodial” nodes) of a given failed span. Thus, integrating node failure recovery in span-protecting networks is much more challenging than with path-oriented protection.

If bidirectional-line-switched-rings (BLSRs) and unidirectional-path-switched rings (UPSRs) are

recognized an inherent ability to recover paths transiting through a failed node, within the surviving portion of the ring, node failure protection using span-oriented paradigms in mesh-based survivable networks requires (in fact) some extensions of the original principle. A related illustration is the node-inclusive span survivability (NISS) scheme for span restoration in [2]. The key idea behind NISS is to define two custodial regions, one hop away from custodial nodes with respect to each possible span failure. Doing so, a related, relatively small and localized instance of the path restoration problem can be solved for any failure affecting each given node-inclusive span entity (i.e., each span plus its custodial nodes).

p-Cycles are now a fairly known span-protecting scheme, with many interesting and attractive properties [3]-[5]. The original intention with span-protecting or “ordinary” p-cycles (as opposed to more recent FIPP p-cycles [13]-[18]) is efficient and fast protection against single span failures. A subsequent and common misunderstanding is that span-protecting p-cycles offer no form of node protection. More correctly, since inception, it has been realized that p-cycles do offer inherently the same protection to on-cycle paths traversing a failed node, as does a BLSR with respect to paths in the ring [6]-[7].

What has, however, remained less clear is how to protect paths that transit a node on a p-cycle and which have straddling relationship to the respective p-cycle. To protect those straddling paths against node failures as well, there have been various extensions to the basic node-protecting property of p-cycles. One main idea explored for node protection with p-cycles is the “node-encircling” principle studied and developed in, for example, ([6],[8]-[9]). Another line of work partly motivated by including node protection has lead to extensions of the whole p-cycle concept into path-segment or so-called “flow-protecting” p-cycles [10]-[11], and to end-to-end path protection with p-cycles [13]-[18].

Overall this contribution is extended from [1]. We explain and explore a two-hop flow strategy to node failure protection using ordinary p-cycles, which seems to have been overlooked to date. Section II presents the two-hop flow concept and compares this with prior related concepts. Section III formulates an integer-linear-programming (ILP) design model for the two-hop flow strategy, recalls equivalent ILP mathematical models for prior approaches, and proposes an adaptation of a novel combination of genetic-algorithms (GA) and ILP methods ([18]-[20]) to address large scale instances.

Section IV presents case studies, the test methodology and experimental results. Section V concludes the paper and indicates possible lines of future direction.

II. PROBLEM FORMULATION AND RELATED WORK

A. Two-Hop Flow: a New Insight and Approach to Node Failure Recovery using Ordinary p-Cycles

In brief, this contribution is to observe that the BLSR-like loopback reaction that ordinary p-cycles make to restore on-cycle flows transiting through a node is actually also applicable to straddling flows failing at a p-cycle node if the two spans adjacent to the failure node both end on another nodes on the same p-cycle. Of course, in hindsight, this is always true for on-cycle flows transiting a node; so this is a generalization of the prior known BLSR-like node protection condition. But the more general criteria can be seen to also allow protection of the following additional cases (with ordinary p-cycles) shown in Fig. 1.

Fig. 1 actually illustrates how an ordinary span-protecting p-cycle can restore any 2-hop path-segment intersecting the cycle structure upstream and downstream of a given failed node, whether or not the 2-hop flow-segment is entirely on the protecting structure. In Fig. 1(a), we have a p-cycle under the normal network state. In Fig. 1(b)-(e), we show how that p-cycle can be used to react in a previously overlooked way under node failure circumstances. The only requirement is that the end-nodes of the two-hop segment are on the same p-cycle as each other. So, Fig. 1(f) captures the class of situation where at least one end-node of the two hop-segment under consideration is not part of the cycle structure; this cannot be covered by the novel general criterion of node failure recovery.

In Fig. 1(b)-(e), the failure scenarios specifically consider whether the 2-hop segment is entirely on the protecting structure as in Fig. 1(b), or if the p-cycle

crosses only one of the spans of the 2-hop segment as shown in Fig. 1(c), or if both spans of the 2-hop path-segment straddle the cycle as in Fig. 1(d), or if only the end-nodes of the 2-hop flow are part of the cycle structure as shown in Fig. 1(e). One way to think about this is to consider any 2-hop flow as a kind of “virtual span”: doing so the three cases in Fig. 1(b)-(d) are all equivalent to on-cycle (span) failures; while the situation in Fig. 1(e), in which neither of the two spans comprising the 2-hop segment nor the failed node are part of the cycle, corresponds to a p-cycle reacting to a straddling (span) failure.

The practical importance of the 2-hop standpoint is that the simplicity of operation of ordinary p-cycles is retained and only one set of span-protecting candidate structures can be employed in a complete design for both 100% span and node failure protection. Subsequently, the 2-hop strategy stands in contrast with prior attempts to protect against node failure events through concepts such as node-encircling p-cycles (NEPCs—[6],[8]-[9]), full path-segment (or more simply “flow”) protecting p-cycles [10]-[11], and failure-independent path-protecting (FIPP) p-cycles [13] to [18].

B. Two Hops versus Node-Encircling p-Cycles A p-cycle is said to be an NEPC for a given

“encircled” node if it contains all the neighbor-nodes of the encircled node, but not the given node itself. Thus the key property is that, an NEPC intercepts any flow transiting the encircled node, and hence (with suitable capacity) can reroute all affected transiting flows when the node fails. For example, the p-cycle in Fig. 2(a) is an NEPC for node G and, as shown in Fig. 2(b), intercepts every path transiting through the encircled node G. In contrast, the given p-cycle cannot be an NEPC for other vertices of the graph because nodes A, I, D, E, F and H are themselves part of the protecting structure; while off-cycle nodes B and C are neighbors each other and thus, B has its neighbor-nodes C (and C its neighbor node B) out of the cycle structure.

(a) Example of p-

(b) BLSR-like behavior, full on-cycle failed flow

(c) Failed flow partially on the

(d) Failed flow:

on-cycle nodes and straddling spans

(e) Off-cycle node

failure, straddling flow

(f) What cases this criterion does not

cover?

Legend:

p-Cycle structure

Failed 2-hop segment Failed Node Protection segment

Fig. 1 Intrinsic “Two-hop” Node Protecting Capabilities of p-Cycles

But, because an NEPC does not include (by definition) the protected node itself, this approach does not exploit the inherent reaction p-cycles can have against on-cycle node failures. Fairly often as well, some nodes may have no simple NEPCs in a graph, especially in sparser networks. Non-simple candidate cycles crossing a span or node more than once can be still considered, but this adds greatly to the operational and conceptual complexity. Designing a separate set of NEPCs generally also

(a) The p-cycle is an NEPC for G; but is not for B and E

(b) The (blue) NEPC intercepts all flows transiting the encircled node G

Fig. 2 The Concept of NEPCs

requires significantly more spare capacity in the complete design because the NEPCs provide for node recovery separately from other p-cycles which are still needed for span protection.

The 2-hop strategy is more flexible than NEPCs. By considering failed-segments of (only) two hops, we eliminate the node-encircling constraint. Instead, for every potential node failure scenario, we select a subset of ordinary cycles that cover (if considered together) all the neighborhood of the failed node—rather than looking for NEPCs of the failed node. Results to follow show much greater overall capacity efficiency as well.

C. Two Hops versus Flow-Protecting p-Cycles In other prior work, full path-segment (or flow)

protecting p-cycles can also support node failure recovery. The principle is to observe that every p-cycle will also happen to intersect a number of working flows upstream and downstream. Subsequently, any intermediate node or span failure along each respective intersecting flow-segment can be restored within the p-cycle, exactly as with the conventional switching mechanism. For instance, Fig. 3(a) shows two cycles X and Y respectively intersecting path-segments [A,C] and [B,D] of a given working flow. In Fig. 3(b), cycle X offers two restoration routes in the event of span or node failure along segment ]A,C[. This is also the case for cycle Y and the segment-portion ]B,F[; but as shown in Fig. 3(c), there is only one protection route within cycle Y if the span/node failure occurs along the portion [F,D[.

Flow-protecting p-cycles generalize the 2-hop strategy in the sense that protected segments may freely go from one or more spans to entire working paths. But by restricting failed flows to be considered and restored strictly only as if they were two-hop segments, we only require the same simple and local type of failure detection and pre-defined switching plans as for span failures. To illustrate, the situation in Fig. 3(c) requires advanced inter-nodal signaling or centralized management to activate the right restoration actions, which are different depending on where the failed path-segment is disrupted—i.e. whether along ]B,F[ or [F,D[.

However, if the added complexity of a failure-dependent reaction is accepted, we recently showed that very high levels of node failure restorability could be achieved by applying a path-segment view to ordinary span-protecting p-cycles [12].

D. Two Hops vis-à-vis Failure-Independent Path-Protecting p-Cycles

Although this requires switching from span- to path-oriented paradigms, failure-independent path-protecting (FIPP) p-cycles with proper node-disjointness constraints also stand as a valid alternative approach to node failure protection using p-cycles. FIPP p-cycles actually operate like conventional p-cycles but they are chosen so that each protects a set of end-to-end paths that are mutually span- (and when desired, node-) failure disjoint between end-nodes on the FIPP structure [13]-[14]. Literature indicates disjoint-route-sets (DRS) and column-generation (CG) as practical methods for FIPP network planning [15]-[17].

When FIPP failure independency constraint is relaxed so that a given working path can be assigned different p-cycles depending on where the failure occurs, the principle is referred as general path-protecting (GPP) p-cycles. It has been shown in prior research that optimal GPP solutions are very close to being FIPP solutions because in general, no more than 2 working paths remain unprotected after the constraint of failure independence is imposed onto GPP designs [18]. Thus as is the case in this paper, it is not awkward to consider GPPs as if they were FIPPs. The merit is that relative to FIPP p-cycles, GPPs can be very efficiently captured in a mathematical formulation.

III. MATHEMATICAL PROGRAMMING ASPECTS The conventional p-cycle minimum spare capacity

design model is given as starting point.

A. Conventional p-Cycle Minimum Spare Capacity Design Model

The following definitions serve for the p-cycle minimum spare capacity design model.

S is the set of spans in the network, indexed by i for failing spans and j for surviving spans or spans in general.

P is the set of candidate cycles, determined by a pre-processing method and indexed by p.

Input Parameters: jC is the cost of each unit of capacity (i.e. channel)

on span j. is the number of working channels to be

protected on span i. This is an input arising from whatever routing process is employed for demand matrix.

{ }0,1,2pix ∈ encodes the number of restoration

path-segments that a single unit-sized copy of p-

(b) Off-cycle node failure

(a) A working path intersecting two protection cycles

(c) On-cycle node failure

Fig. 3 Flow-Protecting p-Cycles: cycles X and Y handle path-segments [A,C] and [B,D] respectively

cycle p may provide to span i. if i straddles

p, if p crosses i, and otherwise.

2pix =

1pix = 0p

ix = Decision Variables: js is the number of spare channels on span j in the

design. pη is the number of unit-sized copies of p-cycle p in

the design. ILP Formulation: Eq. (1) is to minimize total spare capacity requirements

while Eq. (2) guarantees full protection in the event of single span failures, through the p-cycles built within spare capacity in Eq. (3). Minimize j j

C s∈

⋅∑ . (1)

Subject to: p p

i ip P

w x η∈

≤ ⋅∑ , . (2) i S∀ ∈

s η∈ =

= ∑ , . (3) j S∀ ∈

B. ILP Models for the 2-Hop Approach to Node Failure Protection using Ordinary p-Cycles

The following additional definitions serve for the ILP maximizing R1-node in p-cycle designs.

Additional Sets: N is the set of nodes in the network, indexed by k. D is the set of demands, indexed by r. We assume all

units for a given demand-pair r take the same working route.

Additional Input Parameters: is the number of units of capacity for demand-

pair r.

encodes end-nodes for demand-pair r.

if node k is either origin or destination of r,

otherwise.

{ }0,1rkϑ ∈

1rkϑ =

0rkϑ =

indicates which nodes are on the working

route of demand-pair r; if r crosses k en route,

and otherwise.

{ }0,1rkε ∈

1rkε =

0rkε =

indicates how many protection routes are available within the cycle structure p to restore a 2-hop segment for demand-pair r, which has at its end-nodes on p-cycle p and k as intermediate node: if k is off-cycle; if k is on-

cycle; and otherwise.

{, 0,1, 2p rkμ ∈ }

, 2p rkμ = , 1p r

kμ =, 0p r

Additional Decision Variables: is the number of unit-copies of p-cycle p

allocated to demand-pair r in order to prevent node k failures.

,p rkn

,p rkθ is the number of capacity units for demand-pair

r effectively rerouted within p-cycle p when node k fails.

kΛ , kΤ , kΘ record, in the event of node k failure, statistics on affected, transiting and recovered traffic. The following inequality is always true:

k k kΛ ≥ Τ ≥ Θ . ILP Formulation:

1) Maximizing Node Failure Restorability Level in a Conventionally Designed p-Cycle Network

An assumption to this first new ILP is to keep as is the routing of working paths, the spare capacity and the p-cycles selected in an otherwise conventional p-cycle minimum spare capacity design. And the objective is to maximize node failure restorability as stated in Eq (4).

To calculate node failure recovery level—Eq. (5) allocates restoration path-segments, available within the selected p-cycles, to working paths transiting the intermediate node of any 2-hop segment. Doing so, Eq. (6) asserts that only intersecting flows are potentially restorable. Eq. (7) keeps the assignment of protection path-segments under the actual spare capacity of available p-cycles. And Eq. (8) gives no credit to potentially protected paths that would exceed the actual demand volume present.

Eqs. (9)-(11) are for statistics only: they respectively record the demand volume affected by a given node outage, the amount that is potentially restorable because transiting the failed node, and the number of working paths that are effectively protected in the design. Node

failure restorability is given by 1-node

Τ=∑

Maximize kk N∈

Θ∑ . (4)

Subject to: , , ,p r p r p r

k k kn μ≤ ⋅ r D, ∀ ∈ , , . (5) k N∀ ∈ p P∀ ∈θ, ,p r p r

k kn μ≤ ⋅∞ , r D∀ ∈ , k N∀ ∈ , . (6) p P∀ ∈,p rp

nη∈

≥∑ , k N∀ ∈ , p P∀ ∈ : . (7) 1rkε =

,p r rk

dθ∈

≤∑ , r D∀ ∈ , k N∀ ∈ : and 1rkε = 0k

rϑ = . (8)

r rk k

dε∈

Λ = ⋅∑ , k N∀ ∈ . (9)

r rk k

ε∈ =

Τ = ⋅∑ , k N∀ ∈ . (10)

, : 0kr

p rrk k k

r D p P ϑ

ε θ∈ ∈ =

Θ = ⋅∑ , ∀ ∈ . (11) k N

2) R1-node Maximization with Controlled or No Penalties over Minimum Spare Capacity Requirements

Rather than maximizing node failure recovery in a pre-planned 100% span restorable p-cycle network, we can nudge the minimum spare capacity solution to happen to support simultaneously the maximum feasible level of node failure restorability. Setting a suitable small α, this is achievable with the bi-criteria objective in Eq. (12) and the constraints (2)-(3) and (5)-(11).

Minimize ( )j j kj S k N

C s α∈ ∈

⋅ + ⋅ Τ −Θ∑ ∑ k

. (12)

One might want to assert, instead, maximum R1-node subject to an allowable extra budget ξ relative to the minimum spare capacity cost B for 100% restorability against single span failures. If so, Eq. (13) will be accordingly added to the prior set of constraints and Eq. (4) will be then considered as objective function.

(1j jj S

C s B ξ∈

⋅ ≤ ⋅ +∑ . (13)

In a different way, one can also ask how much spare capacity is required at minimum to guarantee 100% restorability against both node and span failures. The equivalent ILP is given by Eqs. (1)-(11), without the objective function in Eq. (4), and the new constraint (14) that transforms the inequality (8) in an equality satisfying the needs for full node failure recovery.

,p r r rkk

dε∈

Θ = ⋅∑ , , : . (14) r D∀ ∈ k N∀ ∈ 0krϑ =

3) Multiple Quality of Protection Concerns The above-defined mathematical formulations are

easily adaptable to multiple quality-of-protection (multi-QoP) purposes. Assume that a new input parameter

identifies three service classes requiring:

100% R1-node if , maximum possible R1-node within

available p-cycles when , and no form of node

failure protection for .

{0,1, 2r℘ ∈ }2r℘ =

1r℘ =

0r℘ =The multi-QoP design model is defined by the

objective function in Eq. (12) and constraints (2)-(3), (5)-(11) and (14). Doing so, Eqs. (5)-(6) will now apply to the situations where ; and Eqs. 0r℘ > (8) and (14) will

stand for and , respectively. Given that node failure recovery levels are known to be 0 and 100% for the first and third service classes, statistics in Eqs.

1r℘ = 2r℘ =

(9)-(11) will belong to only. 1r℘ =

C. Equivalent ILP Formulations for Prior Node Restoration Options with p-Cycles

1) Flow-Protecting p-Cycles and NEPCs All prior-defined ILPs are applicable as is to the cases

of flow-protecting p-cycles and NEPCs. The principle is just to recognize that each p-cycle has its own spans for which it provides the intended span failure protection but exploits and/or adjusts the design in a way that p-cycles also act as 2-hop protecting p-cycles, flow-protecting p-cycles or NEPCs when it comes to prevent node failures. From the flow-protecting p-cycle perspective, the parameter ,p r

kμ has to be pre-processed for full path-segments (of one or more spans to entire working paths). From the NEPC viewpoint, for every r transiting k if and only if p is an NEPC for node k; otherwise, .

, 2p rkμ =

, 0p rkμ =

On the other hand, because of the scarcity of NEPCs (especially in sparser networks), it might happen that all paths cannot topologically survive every single node failure. In such cases, 100% node (and span) restorable designs are not achievable. Thus, the multi-QoP ILP for NEPCs nudge maximum possible level of node failure restorability for 2r℘ = with the prior bi-criteria objective. Eq. (15) can serve as objective function, with suitable settings for β and γ. And Eqs. (9)-(11) have to be adjusted in a way that Λxk,Τxk and Θxk records statistics

kΛ , kΤ , kΘ for r x℘ = .

Minimize ( ) (2 2 1 1k k j j k kk N j S k N

C sβ γ∈ ∈ ∈

Τ −Θ + ⋅ ⋅ + Τ −Θ )∑ ∑ ∑ .(15)

2) Path-Protecting p-Cycles The following additional definitions serve in GPP

related formulations. Additional Input Parameters: { }0,1r

jδ ∈ indicates spans that demand-pair r

crosses en route; 1rjδ = if r crosses span j, and

0rjδ = otherwise.

{ }0,1, 2pry ∈ encodes the number of protection

segments that one unit-sized copy of p-cycle p may provide to demand-pair r. if working route

of r fully straddles p-cycle p, if r is in a full

or partial on-cycle relationship with p, and

2pry =

1pry =

0pry =

otherwise. Additional Decision Variables: represents the number of copies of p-cycle p

assigned to demand-pair r in the design.

ILP Formulation: The conventional p-cycle minimum spare capacity

design model is not applicable to GPP p-cycles, because the latter requires to switch from span- to path-based protection. To achieve full span restorable GPP designs, Eqs. (1) and (3) can be combined with the following constraints. Eq. (16) guarantees working paths to survive any single span failure. And Eq. (17) selects and capacitate cycle structures, providing enough cycle copies to handle rival working routes that share one or more spans in common.

p p rr r

y m d∈

⋅ ≥∑ , r D∀ ∈ . (16)

p r pr i

m δ η∈

⋅ ≤∑ , i S∀ ∈ , p P∀ ∈ . (17)

Regarding node failure protection purposes, Eq. (18) is an equivalent of Eq. (17) but for node-disjointness requirements. The assignment of protection segments within available p-cycles, previously done by Eqs. (5)-(6), is now achieved by Eq. (19). Eq. (20) calculates cycle copies required in the design, and all Eqs. (7)-(14) remain the same.

p r pr k

m ε η∈

⋅ ≤∑ , k N∀ ∈ , p P∀ ∈ . (18)

, ,p r p r prk kn yθ ≤ ⋅ , , , . (19) r D∀ ∈ k N∀ ∈ p P∀ ∈

,p r prkn m≤ , , , r D∀ ∈ k N∀ ∈ p P∀ ∈ . (20)

Table 1 Summary of ILP Mathematical Design Models

Model Problem Description ILP Formulation 1 Conventional reference

ILP p-cycle minimum spare capacity.

Reference values and actual cycles for use in model 2—Eqs. (1)-(3).

2 Maximum R1-node, given a 100% span restorable design as input.

Assuming an existing set of p-cycles and working paths—Eqs. (4)-(11)

3 p-Cycle minimum spare capacity planning under R1-node maximization.

Bicriteria minimization of capacity and node unrestorability—Eqs. (12), (2)-(3), (5)-(11).

4 R1-node maximization with controlled penalties over minimum capacity.

Merge models 1 and 2 plus extra budget—Eqs. (2)-(11), (13).

5* Full protection against both span and node failures.

Capacity minimization under full R1-node—Eqs. (1)-(3), (5)-(7), (9)-(11), (14).

6 Multi-QoP services for R1-node.

Merge models 3 and 5—Eqs. (2)-(3), (5)-(12), (14).

7 Service differentiation for NEPCs.

Model 6 for NEPCs—Eqs. (2)-(3), (5)-(11), (15).

8 GPP p-cycle minimum spare capacity.

Model 1 for GPP—Eqs. (1), (3), (16)-(17).

9-13** Similar to models 2-6, but for GPP.

Eqs involved—(1), (3)-(4), (7)-(14), (16)-(20).

ILP models 1 to 4 apply to 2-hop, flow and NEPC node failure protection strategies. *Model 5 is for 2-hops and flows only; and has no equivalence for NEPCs.

**Although in vrac, eqs. for ILPs 9 to 13 are picked up similarly to models 2 to 6.

Table 1 gives a summary ILP mathematical design models. All of them were implemented in AMPL 10.100 and solved using CPLEX 10.1.0 with a mixed integer programming gap for optimality (MIPGAP) of 10-4, on an Intel Duo Core Processor running Mac OS X 10.5.8 at 2.8 GHz with 4 GB of 1067 MHz DDR3. Where ILP problem instances were solvable with the complete set of candidate cycles, the whole process (including preparatory programs) typically reached full termination in about 15 minutes or less.

We managed the size of large problem instances by restraining the number of eligible cycles. Our preselection technique relates to a novel combination of GA with ILP methods which seems to have many features to recommend it for any large p-cycle problem involving the selection of a relatively few optimal candidate cycles from an almost infinite space. Note that this added (only) a couple more minutes in running times for network instances under consideration in this paper.

D. A GA-ILP Heuristic to Solving Node-Protecting p-Cycle ILPs at Very Large Scale

The GA-ILP preselection concept follows the normal steps of a GA-like evolutionary heuristic. The initial set of candidates is first partitioned into subsets of equal sizes, each subset comprising an individual of which genome corresponds to the index numbers of the cycles constituting the subset in question. The union of the above individuals then embodies an initial population for the GA-ILP, and the node-protecting p-cycle problem under consideration is solved using in turn each of the individuals. Every individual is subsequently assigned a weight equivalent to the objective function value of the constituent ILP and the n/2 individual-pairs showing the optimum sum of weights are selected for breeding. Every selected pair of individuals produces two children by crossing the first half of one parent’s genome with the second half of the second individual’s genome and vice versa. To maintain genetic diversity into the offspring, there is a specific mutation policy that consists of randomly substituting cycle indexes in some children for solution cycles of the individuals not selected to reproduce. The generational process of evaluation, selection, crossover and mutation is repeated on new populations, until the objective function value of the constituent ILP does not improve anymore from one individual to another. All unique cycles of the last population then comprise the reduced set of candidates that is still, to our knowledge, O(individual size).

Whereas the space of all possible candidates is fully enumerable, the GA-ILP solution to a conventional p-cycle network design problem is expected to be equivalent to what would be obtained if the instance under consideration was solved with the entire set of candidate cycles. As a form of test, in work presented at [21] we showed that even for almost 85,000 candidate structures, the GA-ILP always reaches optimality for

problem instances to which exact (ILP) solutions are known. But the purpose for GA-ILP is to go onto much larger problem sizes where enumeration is not practically possible. In this regard, we can actually recognize [21] two additional problem classes based on whether the space of all possible candidate cycles is enumerable but impractical to import into the ILP solver, or not even enumerable in practice. The GA-ILP provided a high quality solution for an instance of about 387,740 candidates. In fact, that solution was found equal (within the MIPGAP being employed) to what was obtained using the CG approach, meaning that the GA-ILP was still within 1% of optimality. In other prior work, we successfully applied the GA-ILP framework for p-cycle network designs with controlled optical path length in the restored network state [19], for near-optimal FIPP p-cycle network designs through GPP [18], and to a 200-node challenge case which represents a specific instance of the third class of problems [20].

IV. EXPERIMENTAL RESULTS AND DISCUSSION

A. Case Studies Five test case networks, shown in Fig. 4, were used.

The first group of columns in Table 2 gives their number of nodes |N|, spans |S|, demand-pairs |D| and eligible cycles |P|. “Havana” in Fig. 4(a) is a previously used network [22]. Two sets of demands are considered for it: the original matrix of 58 demand-pairs with units distributed on the interval [0..5], and another traffic matrix assuming connection requests between every single pair of nodes, with volumes uniformly distributed

on the interval [1..100]. The 2nd network is the well-known Cost239 pan-European network, in Fig. 4(b). Related traffic matrix includes 55 non-zero demand-pairs uniformly distributed on [1..10]. The other test case instances are given in Fig. 4(c)-(e): the Italy network has 78 demand-pairs distributed on [1..10], the Bellcore network has 104 demand-pairs distributed on [1..20], and another European network has 323 demand-pairs distributed on the interval [1..2].

A shortest distance based routing is applied under most normal network states, resulting in total working channel-kms of 23,934 for Havana when considering the original set of demands; 2,595,800 for Havana using the large traffic matrix; 137,170 for Cost 239; 62,232 for Italy; 18,335.1 for Bellcore; and 246,375 for Euro network. There is one more Havana network instance in which a hop-count based routing is applied to the original demand matrix; this leads to 166 working channels. The 3rd and

4th columns report statistics on paths affected by potential node failure events and the number of such paths that can be considered for restoration—the ratio of transiting over total paths (including terminating paths which cannot be restored) varies from 25 to 60%.

Table 2 Test Case Networks Characteristics Networks |N| |S| |D| |P| Λ Τ

17 26 58 135 Havana 32 nepcs, cover 4 nodes

(i) Initial routing

Shortest distance based routing → 23,934 working channel-kms.

271 77 i.e. 28% of affected

(ii) Other routing

Least hop-count based routing → 166 working units.

(iii) Larger traffic matrix

136|D| uniformly distributed on [10..100] → 2,595,800 working channel-kms.

29,317 14,591 50% of affected

11 26 55 3531 Cost239 (shortest distance routing)

- Total of 137,170 working channel-kms. - 1735 nepcs, cover all nodes.

13 24 78 557 Italy (shortest distance routing)

- Total of 62,232 working channel-kms. - 359 nepcs, cover 11 nodes.

15 28 104 976 Bellcore (shortest distance routing)

- Total of 18,335.1 working channel-kms. - 847 nepcs, cover 8 nodes.

32 42 323 699 Euro (shortest distance routing)

- Total of 246,375 working channel-kms. - 458 nepcs, cover 9 nodes.

2483 1497 ie. 60% of affected

Six types of results are considered here to assess effectiveness of the proposed “two-hop” node recovery principle. In a first set of experiments corresponds to models 1 and 8, in which nothing special is done for node failure recovery in the design. These are test cases designed for R1-span= 1 at minimum spare capacity. Within each network, we then use the 2nd and 9th ILPs to simulate each node failure and experimentally determine what is the best R1-node level that can be obtained through the two-hop and other comparative node recovery methods. Further types of result obtained with models 3 and 10 show the level of R1-node that is achievable “for free” under each principle (i.e., with no investment beyond that needed only for R1-span= 1; but free to bias the solution towards choosing cycles that also increase R1-node levels). With enhanced ILPs 4 and 11, we maximize node failure protection under given spare capacity budgets. The 5th and 12th ILP-based results are to see how much spare capacity has to be added to strictly assert 100% R1-

node by each method being compared. The final type of results relates to service differentiation and is achievable with the 6th, 7th and 13th ILPs.

B. Performance of the 2-Hop Strategy The 2nd column of Table 3 gives conventional p-cycle

minimum spare capacity design solutions for each of the networks under consideration; these are 100% span restorable only, with no node failure concerns. The 3rd and 4th columns in Table 3 characterize node failure protection aspects, using each of the different node restoration options. More specifically, the 3rd columns show R1-node levels that are achievable in networks designed for minimum spare capacity but with enhanced R1-node in mind. And the 4th columns report the amounts of added capacity (over min-costs) required to reach 100% restorability against both single span and node failures. Let us first consider the 2-hop flow performance in the first series of data in the 3rd and 4th columns of Table 3.

COPCOP

BERBER

PRAPRA

VIEVIE

MILMIL

ZURZUR

PARPAR

LONLONAMSAMS

LUXLUXBRUBRU

(b) Cost239: 11 nodes, 26 spans, 3531

candidates

N10 N9

N2N1 (d) Bellcore: 15 nodes, 28 spans, 976

candidates

BERHAN

DORESS

KOLLEI

NURMAN

KARSTU

ULMMUN

(a) Havana: 17 nodes, 26 spans, 135

candidates

BREBRE

MILMIL

MITMIT

ALEALE

TORTOR

VICVIC

VENVEN

SAVSAV

GENGEN BOLBOL

PISPIS FIRFIR

VERVER

(c) Italy: 13 nodes, 24 spans, 557 candidates

COPCOP

BERBER

PRAPRA

MUNMUN

MILMIL

STOSTO

WARWAR

BUDBUD

VIEVIE

VENVENTORTOR

ROMROM

LYOLYO

MARMAR

BARBAR

NARNAR

BORBOR

BILBIL

MADMAD

VALVAL

LUXLUX

HAMHAM

DUSDUS

FRAFRA

BRUBRU

AMSAMS

STRSTR STUSTU

ZURZURGENGEN

PARPAR

LONLON

(e) Euro: 32 nodes, 42 spans, 699 candidates

Fig. 4 Five Test Case Networks

Overall, very high levels of R1-node, typically 77% to 96% are achieved under min-costs. Moreover, with additional capacity penalties of 0.89 to 21% we can achieve full node restorability, using the ILP for 100% R1-node.

Fig. 5 gives a deeper analysis of the “Havana” network with the original traffic matrix of 58 demand-pairs over shortest distance routes. The x-axis indicates each node failure scenario and the 1st set of data (i.e., blue histogram) records the corresponding numbers of affected paths. Frankfurt and Hanover have the most impact on the network as each failure affects 37 working paths. In contrast, failures of Bremen and Norden have the least impact on the network, with 4 affected paths each. Overall, there are 271 combinations of node failures and affected working paths. The second histogram bars (i.e., red) in Fig. 5 indicate the number of paths that are potentially restorable because they are “transiting” through a failed node. For example, Dortmund failure affects 26 working paths, but only 9 of those can be considered for restoration because the 17 others are terminating demands at that node. The blanks in Berlin, Hamburg, Munchen and Norden arise because none of the failed paths are transiting those nodes. In totality, 77

failed paths are potentially restorable. The 3rd set of data (i.e., yellow) in Fig. 5 reports maximum achievable levels of single node failure restorability in the conventional design, initially planned for 100% span failure restorability only with no node failure concerns. All working paths transiting failed nodes other than Bremen and Hanover are fully restorable. Bremen and the degree-6 Hanover are respectively 50% and 82% node restorable. Overall, a total of 74 paths (out of 77) survive single node failure conditions, for a very high level of up to 96% R1-node. And this can be pushed to 100%, using the 5th ILP for assertion of 100% R1-node, with less than 1% of additional spare capacity requirements. (The conventional p-cycle minimum spare capacity design solution for Havana involves 16 channel-copies of 4 distinct p-cycle structures, for spare capacity requirements of 20,264 channel-kms corresponding to 84.56% of redundancy to total distance-weighted working capacity.)

BERBRE

DORDUS

ESSFRA

HAMHAN

KARKOL LE

MUNNOR

NURULM

Single Node Failure Scenarios

271 Affected Paths 77 Transiting Paths, i.e. 28.41% 74 Surviving Paths, i.e. 96.10%

Fig. 5 Havana—Node Payload under Normal Conditions vs. Paths Surviving Failures

C. Comparison to Prior Related Approaches Let us now compare the 2-hop approach to the three

other approaches; i.e. NEPC, flow-protecting p-cycles and GPP, for which results are given by 2nd, 3rd and 4th sets of data in the 3rd and 4th columns of Table 3. Noticeably the 2-hops approach is nearly as capacity-efficient as path-segment protecting p-cycles, with a difference of 0 to 20% for both metrics under observation. With 2-hops, flow-segments are otherwise shortened to guarantee, in addition, a straightforward failure detection and a real-time activation of right restoration processes.

NEPCs show R1-node is almost non-existent under min-costs; furthermore, it is usually not topologically possible to achieve full R1-node, and it is too costly when possible as with the Cost239 network. Consider again the Havana network as a supporting example. Table 2 indicates 32 NEPCs over the 135 distinct simple candidates available across the network graph. Those NEPCs cover only four nodes, i.e., Berlin, Hamburg, Hanover and Norden;

Table 3 Sample Results on ILPs 1-4 and 8-11 Surviving Paths under Min. Capa.

Design, Level of R1-node

Min. Capa. for 100% (or max) R1-node Networks Conventional Design (spare capacity,

redundancy) 2-hop nepc flow gpp 2-hop nepc flow gppHavana (i) Initial routing

20,264 i.e. 85% (gpp: 20,451 i.e. ∆=+0.92%)

74 i.e. 96%

0 74 i.e. 96%

77 i.e. 100%

20,444 i.e. ∆=+0.89%

11 i.e. 14%, for 28,134 ∆=+39%

20,335 i.e. ∆=+0.3%

20,451 i.e. ∆=+0.92%

(ii) Other routing

134 i.e. 81% (gpp: 157 i.e. ∆=+17.16%)

53 i.e. 77%

0 53 i.e. 77%

58 i.e. 84%

186 i.e. ∆=+39%

17 i.e. 25% for 237 ∆=+77%

180 i.e. ∆=+34%

197 i.e. ∆=+47%

(iii) Larger traffic matrix

3,187,827 i.e. 123% (gpp: 3,174,532 i.e. ∆=-0.42%)

12,709 i.e. 87%

0 13,177 i.e. 90%

12,860 i.e. 88%

3,875,460 i.e. ∆= +21%

1889 i.e. 26% for 3,550,540 ∆=+11%

3,480,250 i.e. ∆=+9%

3,914,447 i.e. ∆=+23%

Cost239

85,640 i.e. 62% (gpp: 75,970 i.e. ∆=-11.29%)

92 i.e. 77%

2 i.e. 2%

94 i.e. 79%

98 i.e. 82%

99,850 i.e. ∆= +17%

(full) 203,720 ∆=+138%

98,280 i.e. ∆= +15%

79,705 i.e. ∆=-7%

55,654 i.e. 89% (gpp: 47,735 i.e. ∆=-14.28%)

468 i.e. 90%

2 i.e. 0.4%

468 i.e. 90%

457 i.e. 88%

65,696 i.e. ∆= +18%

359 i.e. 69% for 96,518 ∆=+73%

65,489 i.e. ∆= +18%

73,093 i.e. ∆=+31%

Bellcore

14,591.5 i.e. 79.58% (gpp: 13,808.03 i.e. ∆=-5.37%)

338 i.e. 85%

5 i.e. 1%

338 i.e. 85%

329 i.e. 83%

16,199.3 i.e. ∆= +11%

135 i.e. 34% for 24,329.4 ∆=+67%

16,059.1 i.e. ∆= +10%

14,537.1 i.e. ∆=-0.37%

235,206.5 i.e. 95.46% (gpp: 213,532.3 i.e. ∆=-9.21%)

1287 i.e. 86%

83 i.e. 6%

1350 i.e. 90%

1272 i.e. 85%

257,598.81 i.e. ∆=+9.52%

415 i.e. 28% for 388,826 i.e. ∆=+65%

252,715.4 i.e. Δ=+7%

223,872.1 i.e. Δ=-5%

1200 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85

Additional Capacity over Minimum Cost Requirements (%)

2-hop flow nepc

(a) Havana

0 10 20 30 50 70 90 110

Additional Costs over Minimum Spare Capacity Requirements (%)

flow 2-hop nepc

(b) Cost239

Fig. 6 R1-node Maximization under Controlled Penalties over Minimum Spare Capacity

among them, sole Hanover was carrying (about 11) transiting paths in the node payload characterization previously shown in Fig. 5. The 11 working paths thus correspond to the maximum achievable R1-node of 14%, which requires about 39% of extra spare capacity over min-costs. The usage of non-simple cycles is necessary to achieve an R1-node of 100%. In contrast to NEPCs, the path-segment strategy greatly improves on both maximum R1-node under min-costs and spare capacity penalties to reach 100% R1-node. (Another bench of comparison, based on the 4th ILP, is given in Fig. 6 for Havana and Cost239 network; y-axis maximizes R1-node under extra spare capacity indicated in x-axis).

GPP involves shifting from span to path protection, so the related minimum spare capacity design solutions differ from that of conventional span-protecting p-cycles (otherwise used as benchmark for 2-hops, flows and NEPCs). Comparative min-cost solutions are both given in the 2nd column of Table 3. Overall, results seem to suggest that FIPP gives rise to least costs for instance cases involving more candidate cycles while conventional p-cycles are more adapted to smaller candidate spaces. For example, the Havana minimum capacity GPP-design solution is 100% R1-node; there is however an indirect penalty as GPP minimum spare capacity design (i.e. 20,451 channel-kms) is more expensive than that of ordinary p-cycles (i.e. 20,264 channel-kms). Furthermore, in these test networks at least, GPP min-cost requirements

are even higher than what is required to reach full node restorability using either flows (i.e., 20,335) or 2-hops (i.e., 20,444).

D. Performance under Multi-QoP Requirements Overall, the 2-hop strategy shows such a great

performance that it seems to be a promising mixed priority service environments. Table 4 records multi-QoP experimental results for Havana network with the original traffic matrix and the shortest distance routing. As shown in the first row, four scenario types were considered: no node failure protection, maximum failure protection under minimum spare capacity requirements, 100% node failure restorability and several mixed scenarios.

To distribute traffic among service classes for the scenario (50, 30, 20), for example, we randomly generated a number on the interval [1..100], considering in turn each demand-pair. Paths of the respective demand were then considered from class 2 (i.e. 2r℘ = ) if the generated number was on the interval [1..50], class 1 if on [51..80] and class 0 if in [81..100].

Results in the 2nd row in Table 4 indicates same capacities as in the conventional minimum spare capacity design for almost all multi-QoP scenarios. Up to 50% of traffic flows can be offered full node failure recovery while maintaining more than 90% of R1-node for 3/5th of the remaining demands, with exactly zero penalty over

Table 4 Sample Results on Multi-QoP

scenarios (0,0,100) (0,100,0) (0,0,100) (50,30,20) (50,20,30) (20,50,30) (30,50,20) (30,20,50) (20,30,50) spare capa redundancy

20,264 i.e. 84.66%

20,444 i.e. 85.42%

20,264 i.e. 84.66%

affected (0, 0, 271) (0, 271, 0) (271, 0, 0) (141,104,26) (141,65,65) (52,154,65) (76,169,26) (76,65,130) (52,89,130) transiting (0, 0, 77) (0, 77, 0) (77, 0, 0) (39, 28, 10) (39, 21, 17) (12, 48, 17) (22, 45, 10) (22, 17, 38) (12, 27, 38) protected (0, 0, n/a) (0, 74, n/a) (77, 0, n/a) (39, 26, n/a) (39, 19, n/a) (12, 46, n/a) (22, 43, n/a) (22, 17, n/a) (12, 27, n/a) R1-node for

1r℘ =n/a 96.10% n/a 92.85% 90.47% 95.83% 95.56% 100% 100%

what is required for 100% R1-span designs. On the other hand, more than 96% of R1-node is achievable if all demands are considered from class 1 (i.e. 1r℘ = ). R1-

node=1 is possible for every traffic flow with less than 1% of extra spare capacity. (The 3rd, 4th and 5th rows give statistics on affected, transiting and protected paths in the event of any node failure.)

E. Effectiveness of the GA-ILP In experiments conducted, ILP models 3 to 7 were too

large for practical solution using the complete set of candidate cycles in Cost239 and Euro graphs. And we faced this problem more often with GPP ILPs 9 to 13, even when conventionally designed Cost239, Bellcore and Euro networks. Large problem instances were addressed within the GA-ILP framework in Section III.D. Fig. 7 shows the GA-ILP convergence for the 2-hop variant of model 5, i.e. 100% R1-node. In related experiments, partitions of the complete space of candidates correspond to n=6 individuals for Havana, 40 for Cost239, 20 for Italy and Euro networks. The generational process typically completed earlier before the 20th iteration. For Havana, Italy and Bellcore networks of which exact solutions were known, the GA-ILP solution was always within 1% of optimality. Thus, we trusted the GA-ILP solutions for the Cost239 and Euro networks. (We reached the same conclusions for Cost239, Euro and Bellcore when substituting the constituent p-cycle model Euro for flow, NEPC and GPP approaches.)

V. CONCLUSION Recent work has revealed a new, relatively simple and

possibly cost-effective approach to achieve combined protection of optical networks against both node and span failures. The new principle is based on a generalization of how nodes in a BLSR-ring or p-cycle (to date) derive survivability through loopback at the nearest two neighbor-nodes on the same ring. The generalization views any combination of node failure and an affected

transiting path from the standpoint of the 2-hop segment defined by the failure node, and the nodes immediately adjacent on the affected path. We then ask whether these nodes are found together within the same p-cycle as the failure node, or another p-cycle entirely. In any case where they are, we show that the transiting path affected by the node failure is inherently restorable by ordinary p-cycle switching actions whether the respective two-hop segment is on-cycle, straddling, or partially on-cycle and partially straddling. The novel combination of GA-methods with ILP was adapted for node-protecting p-cycles through 2-hop, flow and NEPCs.

The resulting network designs use only a single set of p-cycle structures that have the same or only slightly more capacity than a corresponding optimal set of p-cycles for span protection “only”. In this paper, we explained the principle and characterize its effectiveness in terms of network-wide single node failure restorability (R1-node) in networks designed only for minimum spare capacity, networks designed for enhanced R1-node (at min capacity) and networks designed strictly for R1-node=1. As well, the proposed approach for node-protecting p-cycles is compared to related prior concepts. A subsequent line of future direction is to develop 2-hop protecting p-cycles towards no more distinction between node and span failures.

REFERENCES

[1] W.D. Grover and D.P. Onguetou, “A New Approach to Node-Failure Protection with Span-Protecting p-Cycles,” invited paper for the 11th Intl. Conf. on Transparent Optical Networks—5th Workshop on Reliability Issues in Next Generation Optical Networks (ICTON/RONEXT), Island of Sao Miguel (Azores), Portugal, Jun. 2009.

[2] J. Doucette and W.D. Grover, “Node-inclusive span survivability in an optical mesh transport network,” 19th annual National Fiber Optics Engineers Conf. (NFOEC), pp. 634-643, Orlando, Sep. 2003.

[3] W.D. Grover, Mesh-Based Survivable Networks: Options and Strategies for Optical, MPLS, SONET and ATM Networking, PTR Prentice-Hall, 2003, Chap. 10.

[4] W.D. Grover and D. Stamatelakis, “Cycle-oriented distributed preconfiguration: ring-like speed with mesh-like capacity for self-planning network restoration,” IEEE Intl. Conf. on Communications (ICC), Atlanta (Georgia), USA, Jun. 1998.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Iterations

lty in

Havana Cost239 Italy Bellcore Euro

Fig. 7 GA-ILP Solutions for Full R1-node with 2-Hops

[5] D. Stamatelakis and W.D. Grover, “Theoretical underpinnings for the efficiency of restorable networks using pre-configured cycles (“p-cycles”),” IEEE Transactions on Communications, vol.48, no.8, pp. 1262-1265, Aug. 2000.

[6] D. Stamatelakis and W.D. Grover, “IP layer restoration and network planning based on virtual protection cycles,” Journal on Selected Areas in Communications (JSAC)—Special Issue on Protocols and Architectures for Next Generation Networks, vol. 18, no. 10, pp. 1912-1913, Oct. 2000.

[7] D. Schupke, “Automatic protection switching for p-cycles in WDM networks,” Optical Switching Networking (OSN), Elsevier, vol. 2, no. 1, pp. 35-48, May 2005.

[8] J. Doucette, P. Giese and W.D. Grover, “Combined node and span protection strategies with node-encircling p-

cycles,” Design of Reliable Communication Networks (DRCN), pp. 213-221, Ishia (Naples), Italy, Oct. 2005.

[9] J. Doucette, W.D. Grover and P. Giese, “Physical-layer p-cycles adapted for router-level node protection: a multi-layer design and operation strategy,” Journal of Selected Area in Communications (JSAC), vol. 25, no. 5, pp. 963-973, Jun. 2007.

[10] G. Shen and W.D. Grover, “Extending the p-cycle concept to path-segment protection,” IEEE Intl. Conf. on Communications (ICC), pp. 1314-1319, Anchorage, 11-15 May 2003.

[11] G. Shen and W.D. Grover, “Extending the p-cycle concept to path-segment protection for span and node failure recovery,” Journal of Selected Area in Communications (JSAC), vol. 21, no. 8, pp. 1306-1319, Oct. 2003.

[12] D.P. Onguetou and W.D. Grover, “A new insight and approach to node failure protection with ordinary p-cycles,” IEEE Intl. Conf. on Communications (ICC), pp. 5145-5149, Beijing, China, May 2008.

[13] W.D. Grover and A. Kodian, “Failure-independent path protection with p-cycles: efficient, fast and simple protection for transparent optical networks,” Intl. Conf. on Transparent Optical Networks (ICTON), Barcelona, Spain, Jul. 2005.

[14] A. Kodian and W.D. Grover, “Failure independent path-protecting p-cycles: efficient and fully pre-connected optical-path protection,” Journal of Lightwave Technology (JLT), vol. 23, no. 10, Oct. 2005.

[15] A. Kodian, W.D. Grover and J. Doucette, “A disjoint route sets approach to design of failure-independent path-protecting p-cycle networks,” Design of Reliable Communication Networks (DRCN), pp. 231-238, Ischia (Naples), Italy, Oct. 2005.

[16] D. Baloukov, W.D. Grover and A. Kodian, “Towards jointly optimized design of failure independent path protecting p-cycle networks,” OSA Journal of Optical Networking, vol.6, no. 12, pp. 62-79, Dec. 2007.

[17] B. Jaumard, C. Rocha, D. Baloukov and W.D. Grover, “A column generation approach for design of networks using path-protecting p-cycles,” Design of Reliable Communication Networks (DRCN), La Rochelle, France, Oct. 2007, pp. 243-250.

[18] D.P. Onguetou, D. Baloukov and W.D. Grover, “Near-optimal FIPP p-cycle network designs using general path-protecting p-cycles and combined GA-ILP methods,” Design of Reliable Communication Networks (DRCN), Washington DC, USA, Oct. 2009.

[19] D.P. Onguetou and W.D. Grover, “Approaches to p-cycle network design with controlled optical path lengths in the restored network state,” OSA Journal of Optical Networking (JoN), vol. 7, No 7, Jul. 2008, pp. 673-691.

[20] D.P. Onguetou and W.D. Grover, “Solution of a 200-node p-cycle network design problem with GA-based pre-selection of candidate structures,” IEEE Intl. Conf. on Communications (ICC), Dresden, Germany, Jun. 2009.

[21] W.D. Grover and D.P. Onguetou, “Towards solution of very large p-cycle network design problems with combined GA-ILP heuristic methods,” Optimization of Optical Networks (OON), Montréal, Canada, May 2008.

[22] A. Grue, W.D. Grover, B. Forst, D.P. Onguetou, D.Baloukov, J.Doucette, M.Clouqueur and D.Schupke, “Comparative study of fully pre-cross-connected protection architectures for transparent optical networks,” Design of Reliable Communication Networks (DRCN), La Rochelle, France, Oct. 2007.

Diane Prisca Onguetou is a PhD candidate in Electrical and Computer Engineering department at TRLabs and the

University of Alberta, under the direct supervision of Professor Wayne Grover. She successfully completed in 2005 a M.Sc. in Telecommunications at INRS, Montréal (Québec), Canada. In 2002, she had a special postgraduate training in Signal Processing at "Université de Marne-La-Vallée" in France, and in 2001 she obtained an

Engineering Diploma in Electrical and Telecommunications from "Ecole Polytechnique de Yaoundé", Cameroon. Her PhD research project seeks the goal of increasing the awareness and understanding of p-cycle design methods, through an ongoing series of advanced research in the design of p-cycle survivable networks. Findings to date include an elegant control of restored state path lengths, a cost-effective generalization of how nodes in a BLSR-ring or p-cycles derive survivability through loopback nearest two neighbor nodes on the same ring, whole fiber switched p-cycles and a novel combination of ILP and GA methods to solve the large scale p-cycle design problem. Her PhD defense is expected in Summer 2010. Wayne D. Grover holds a B.Eng. (EE) from Carleton University, Ottawa; an M.Sc.(EE Science) from the University

of Essex, England; and a Ph.D. (EE) from the University of Alberta. Dr. Grover has issued patents on 26 topics each issued in several countries, 76 journal publications, six book chapters and over 100 technical reports, seminars, and conference papers. In August 2003 his book Mesh-based Survivable Networks: Options and

Strategies for Optical, MPLS, SONET and ATM Networking was published by Prentice-Hall (841 pages plus web-based appendices). Three of his research papers, and his Ph.D. thesis on Self-Healing Networks, have become "highly cited" in different technical areas but he is most widely recognized for work in restorable network design and operation, including Sonet, ATM, DWDM and IP/MPLS networks. Following his decade-long development and advocacy of the concepts of self-healing and self-organizing transport networks, he is considered as a founding inventor in this field. In 1999 he received the IEEE Baker Prize Paper Award for his paper “Self-organizing Broadband transport networks” in the Oct. 1997 IEEE Proceedings. Other research contributions are in the areas of high-speed synchronization, precise time transfer, wireless traffic analysis, rate-adaptive subscriber loops, radio-location in wireless systems and availability analysis of transport networks. Dr. Grover was an NSERC E.W.R. Steacie Fellow for 2001-2002. Previously he was the McCalla Professor in Engineering and recipient of the Martha Cook-Piper Research Prize (both at U.of A.), the "Smart City" Award (City of Edmonton) and a Technology Commercialization Award from TRLabs (1997) for the licensing of technology to industry. In 2002 he was made an IEEE Fellow (“for contributions to survivable and self-organizing broadband transport networks),” and Fellow of the Engineering Institute of Canada. In 2003 he is serving as General Chair for the 4th International Workshop on Design of Reliable Networks (DRCN 2003), Banff, Alberta, Oct. 19-22,2003.

Impairment Aware RWA in Optical Networks:

Over-provisioning or Cross Optimization?

K. Christodoulopoulos, P. Kokkinos, K. Manousakis, E. A. VarvarigosComputer Engineering and Informatics Department, University of Patras, Greece, and

Research Academic Computer Technology Institute, Patras, Greece

Email: {kchristodou, kokkinop, manousak, manos}@ceid.upatras.gr

Abstract—In transparent and translucent wavelength routed

optical networks the signal quality degrades due to physicallayer impairments while the interference among lightpathsimplies that routing decisions for one lightpath affect and

are affected by the decisions made for other lightpaths. Toestablish a lightpath for a new connection two main

approaches can be used. The most common approach is toselect a lightpath that has acceptable transmission quality

under a worst case interference assumption, ensuring that

the selected lightpath will not become infeasible due to thepossible establishment of future interfering connections.

This approach sacrifices candidate path space for a quickand stable lightpath selection, which is appealing from a

complexity viewpoint. The second approach is to considerthe current network utilization and account for the actualinterference among lightpaths, performing a cross layer

optimization between the network and physical layers. In

this case, however, the algorithm has to check whether the

establishment of the new lightpath turns infeasible some ofthe already established connections. The question that arisesis whether the performance benefits that can be achieved

through the second approach are worth the addedcomplexity introduced by the cross-layer optimization

applied.

Index Terms— Routing and Wavelength Assignment,

transparent networks, translucent networks, physical layerimpairments, network provisioning

I. INTRODUCTION

In opaque networks the signal is regenerated at every

intermediate node along a lightpath via Opto-Electro-

Optical (OEO) conversion. The network cost could be

reduced by employing regenerators only at specific nodesof the network. When regenerators are available, a

lengthy end-to-end connection that needs regeneration at

some intermediate node(s) is set up in a multi-segment

manner so that it is served by two or more consecutive

transparent lightpath-segments. Optical networks, where

some lightpaths are routed transparently, while others go

through a number of regenerators, are known as

translucent optical networks. In some networks it is also

feasible for the data signal to remain in the optical

domain for the entire path and these networks are known

as transparent networks.

In transparent and translucent networks, it is importantto propose algorithms that select the routes for the

connection requests and the wavelengths that will be used

on each of the links along these routes, so as to optimize

certain desired performance metrics. This is known as therouting and wavelength assignment (RWA) problem. An

offline RWA algorithm is executed when the network is

initially set up for network provisioning (i.e., planning

phase of the network), and is also executed periodically,

or when traffic changes substantially. An online RWA

algorithm is executed for new connection requests that

arrive sporadically and have to be served on demand, one

by one (i.e., operational phase of the network).

The typical objectives of the RWA problem are to

reduce both the blocking ratio and the network cost in

terms of Capital Expenditure (CAPEX) and Operational

Expenditure (OPEX). In transparent or translucent optical

networks a connection blocking may occur (i) due to the

unavailability of free wavelengths or links (network-layer

blocking) and (ii) due to the physical layer impairments,

introduced by the non-ideal physical layer, which may

degrade the signal quality to the extent that the lightpath

is infeasible (physical-layer blocking).

Physical layer impairments reduce the number of

candidate paths that can be used for routing. Moreover,

due to certain physical layer effects, routing choices made

for one lightpath affect and are affected by the routing

choices made for the other lightpaths. RWA algorithmsthat take physical layer impairments into account are

referred to as impairment-aware (IA)-RWA algorithms.

There are two approaches to address the IA-RWA

problem while accounting for the interference among

lightpaths. In the first approach, the quality of

transmission (QoT) of a new candidate lightpath is

calculated under the assumption that all wavelengths on

all links are fully utilized. This will be referred to as the

worst case interference assumption. A lightpath chosen in

this way is bound to have acceptable transmission quality

during its entire duration, even if future interferingconnections are established. However, this approach

reduces the candidate path space available for routing,

resulting in larger blocking probability and wasteful use

of network resources. On the other hand, cross-layer

optimization algorithms that use the current network

utilization to estimate the actual interference among

lightpaths are able to explore a larger path space. The

drawback of this second approach is that the IA-RWA

algorithm’s operation becomes more complicated, since

in this case the actual inter-lightpath interference has to

be modeled, and, additionally, the algorithm has to

evaluate if the establishment of a new lightpath will turn

infeasible some of the already established connections.

In what follows we present in more details and

evaluate through simulations typical algorithms that

follow these two approaches. Our aim is to investigate the

performance and tradeoffs involved when using worstcase impairment estimates or actual impairment estimates

in designing and operating an optical network. In

particular, in Section II we report on previous work. In

Section III we give a short description of the physical

layer impairments. In Section IV we outline algorithms

for transparent optical networks that follow the worst

case interference assumption or estimate the actual

interference among lightpaths so as to perform a cross-

layer optimization of the solution between the network

and physical layers. In Section V we describe algorithms

for network provisioning in translucent networksfollowing the two aforementioned approaches.

Simulation results are presented in Section VI, where we

evaluate whether the cross layer optimization algorithms,

for transparent and translucent networks, exhibit

performance benefits that compensate for their increased

complexity. Our conclusions are given in Section VII.

II. PREVIOUS WORK

Different IA-RWA algorithms proposed in the

literature model in different ways the interaction between

the networking and the physical layer, optimizing their

solutions either separately or jointly over these two

layers. Regarding the operation phase of an opticalnetwork, various online IA-RWA algorithms have been

proposed in the literature [2] - [5]. In [2] the authors

decouple the RWA and the IA subproblems, by first

deciding on the lightpath to serve a connection (RWA

subproblem) and then evaluating the feasibility of the

chosen lightpath on a separate module (IA subproblem).

In [3] an IA-RWA algorithm that selects a lightpath and

then uses analytical models to estimate its QoT is

presented. An IA-RWA algorithm that is based on the

shortest path or shortest widest path concept and uses

analytical formulas to estimate the QoT of each candidatelightpath is presented in [4].

The multicost algorithm presented in our previous

study [5] solves the IA-RWA problem jointly and takes

into account the interference among the lightpaths, using

the current network utilization. This is done by first

calculating noise variance vectors per wavelength that are

used as cost vectors for the links of the network. It then

calculates the path parameter vectors by using appropriate

associative operators to combine the corresponding link

parameter vectors. During its operation the algorithm

calculates the Q-factor of candidate lightpaths and prunes

those that do not have acceptable QoT. In the end, itobtains a set of non-dominated paths from source to

destination that all have acceptable QoT performance.

We turn now our attention to translucent networks, as

opposed to transparent networks discussed above. The

majority of RWA algorithms proposed so far for

translucent networks assume a dynamic (online) traffic

scenario. [6] presents a two-dimensional Dijkstra RWA

algorithm for translucent optical networks that assumes a

given placement for the regenerators and a constraint on

the maximum transparent distance. When the length of a

lightpath exceeds a maximum transparent distance bound,

the lightpath is blocked. A different approach for

dynamic resource allocation and routing is considered in

[7] and [8], where spare transceivers (transmitter-receiverpairs or add-drop ports) at the nodes are used to

regenerate signals. This case applies to networks where

the lightpaths initiated and terminated at a node do not

use up all its transceivers, so that some nodes will have

spare transceivers that can be used for regeneration

purposes. A Max-spare algorithm for selecting the

regeneration nodes for a lightpath is proposed in [9] and

compared to a Greedy algorithm used in conjunction with

a wavelength-weighted and a length-weighted RWA

algorithm. In [10], two online RWA algorithms for

translucent networks with sparse regenerator placementare presented. These algorithms assume (i) worst-case

physical penalties (corresponding to a fully loaded

system), or (ii) take into account the current network

status and the actual number of active channels.

In [11], the problem of maximizing the number of

established connections, under a constraint on the

maximum transparent length, is formulated as a mixed-

integer linear program (MILP). Since MILP is NP-hard,

the authors also propose a heuristic algorithm to route

connections. However, [11] does not consider impairment

effects other than the transparent length. A simple

heuristic is given, for placing the fewest suchregenerators to reach a given blocking probability for

dynamic traffic, based on the ranked frequency of

shortest-path routes transiting each node is given in [12].

In [13] the authors address the translucent network design

problem by proposing several regenerator placement

algorithms based on different knowledge of future

network traffic patterns. A quality of transmission-based

heuristic IA-RWA algorithm for translucent networks is

presented in [14]. In the first phase of that algorithm a

random search heuristic RWA algorithm is used and in

the second phase regeneration placement is performedafter estimating the BER of the lightpaths comprising the

solution of the first phase. In our previous study [15], we

examined the offline IA-RWA problem for translucent

networks and proposed an algorithm that selects the

regeneration sites and the number of regenerators that

need to be deployed on these sites for the given set of

requested connections. The problem of regenerator

placement and regenerator assignment is formulated as a

virtual topology design problem.

III. PHYSICAL LAYER IMPAIRMENTS

Several criteria can be used to evaluate the signal

quality of a lightpath. Among a number of measurabletransmission quality attributes, the Q-factor seems to be

more suitable as a metric to be integrated in an RWA

algorithm, because of its immediate relation to the bit

error rate (BER). The Q-factor is the electrical signal-to-

noise ratio at the input of the decision circuit in the

receiver’s terminal [3][4]. Physical layer impairments

(PLIs) are usually categorized to linear and non-linear,

according to their dependence on the power. However,

when we consider IA-RWA algorithms it is useful to

categorize the PLIs to those that affect the same lightpath

(Class 1) and to those that are generated by the

interference among lightpaths (Class 2). Table I presents

this classification.

TABLE I: PHYSICAL LAYER IMPAIRMENTS CLASSIFICATION

Class 1: Impairments that affect

the same lightpath

Class 2: Impairments that are

generated by other lightpaths

Amplified Spontaneous Emission

noise (ASE)

Polarization Mode Dispersion (PMD)

Chromatic Dispersion (CD)

Filter concatenation (FC)

Self-Phase Modulation (SPM)

Intra-channel and inter-channel

Crosstalk (XT)

Cross-Phase Modulation (XPM)

Four Wave Mixing (FWM)

IV. WORST-CASE VERSUS ACTUAL-CASE DESIGN OF

TRANSPARENT NETWORKS IN THE PRESENCE OF PLIS

In this section we discuss ways that can be used to

incorporate physical layer impairments (PLIs) into the

RWA problem. PLIs of Class 1 depend only on the

selected lightpath and can be treated quite easily. Assume

that for a new connection request we can pre-calculate a

set of candidate lightpaths, using some cost criterion. For

each candidate lightpath we can calculate the effects of

the PLIs of Class 1, using, e.g., analytical models, and

discard those with unacceptable QoT performance.

PLIs of Class 2 are more difficult to be accounted for,since they make decisions for one lightpath depend on

decisions made for other lightpaths. These computations,

using analytical formulas, are time consuming for an

online algorithm. Moreover, due to these impairments,

the establishment of future connections may turn

infeasible some previously established lightpaths. An

obvious simplification is to consider a “worst case

scenario”, that is, to assume that all wavelengths in the

network are active, and calculate the worst case

interference accumulated on each candidate lightpath.

Then, the lightpaths that do not have acceptable QoT

performance under this worst case assumption can bediscarded, ensuring that the chosen lightpath is feasible,

irrespectively of the actual utilization of the network.

This approach does not choose the lightpath that is

optimal for the current utilization of the network, but acts

as if the network was fully utilized. In practice, the

wavelength continuity constraint limits the maximum

achievable network utilization, except for the degenerate

case where all connections are between adjacent nodes.

Thus, the key drawback of the worst case interference

assumption is that it results in discarding candidate

lightpaths that are not really infeasible. The actual

feasibility or not of these lightpaths depends on the

lightpaths that are active in the network.

To illustrate this, we quantify through an example case

the degree to which the routing solution space is reduced

when physical layer impairments are considered. We

assume the generic Deutsche Telekom (DTnet) topology,

shown in Figure 1, with physical layer parameters chosen

to have realistic values. We have also used a quality of

transmission evaluation module (Q-Tool) developed

within the DICONET project [16] that uses analytical

models to account for the most important physical layer

effects and in particular all the physical layer

impairments presented in Table I. We assume that there is

a single connection request for each source-destination

pair in the network for a total of N.(N-1) connection

requests, where N is the number of nodes in the network.

For this set of connection requests, we calculate, initially,

k-shortest length paths, for different values of the

parameter k, and then we prune this set of candidate paths

using the Q-Tool by eliminating paths that are estimated

to be infeasible. In doing so, we either assume an empty

network, discarding lightpaths that are infeasible due to

impairments of Class 1, or we assume a fully utilized

network, discarding lightpaths that are infeasible due to

impairments of Class 1 and of Class 2 under the worst

case interference scenario. Table II shows that the pathpopulation obtained after eliminating candidate paths due

to the impairments of Class 1 (column (b)) is

considerably larger than that obtained when we use the

worst case interference assumption for the impairments of

Class 2 (column (c)).

An IA-RWA algorithm that assumes a worst-case

interference and explores the solution space that

corresponds to column (c), is expected to obtain zero

physical-layer blocking, since lightpaths will only be

rejected due to lack of available wavelengths (network-

layer blocking). Moreover, it is guaranteed that the

selected lightpaths will not become infeasible due to the

establishment of future connections. However, such an

algorithm explores a smaller solution space and

unnecessarily restricts the RWA choices, when compared

to an algorithm that takes into account the actual

utilization state of the network and explores the solution

space that corresponds to column (c). This may lead to

deterioration in the performance of the IA-RWA

algorithm that assumes a worst-case interference (higher

network layer blocking). We will come back and quantify

this performance difference later in this article.

A. k-SP worst case IA-RWA algorithm

In this section we outline a simple IA-RWA that

follows the worst case interference approach. We assume

that for each source-destination pair, the algorithm pre-calculates a set of k-shortest length paths (k=5 in our

simulation experiments). Using analytical models for all

TABLE II:THE REDUCTION IN THE SOLUTION SPACE DUE TO PLIS OF CLASS 1 AND

CLASS 2 (UNDER THE WORST CASE INTERFERENCE ASSUMPTION), FOR THE

CASE OF THE GENERIC DT NETWORK TOPOLOGY AND THE REFERENCE

TRAFFIC MATRIX.

(a)Initial path

population

(k-shortest

length paths)

(b)Population after

discarding paths due to

impairments of Class 1

(c)Population after

discarding paths due to

impairments of Class 1

and Class 2 - assuming

worst case interference

k=1 182 182 182

k=2 364 359 333

k=3 546 528 427

k=4 728 653 479

k=5 910 751 506

PLIs described in Section III and under the worst case

interference assumption the algorithm prunes the set of

candidate paths so as to finally keep only the paths that

have acceptable QoT performance (paths belonging to

column (c) of Table II). The current network utilization

state is only considered in order to identify the freewavelengths that are available to serve a new connection.

In particular, when a new connection request for source-

destination pair (s,d) arrives, the algorithm searches the

candidate paths of (s,d) for free wavelengths and selects

from the paths that have at least one available

wavelength, the one that uses the path with the smallest

number of hops, and from the wavelengths of that path,

the wavelength that is utilized most in the network. This

follows the shortest-hop and most used wavelength

approach that is widely used in RWA algorithms [1].

B. k-SP current state IA-RWA algorithm

This algorithm again pre-calculates k-shortest length

paths, but this time analytical models only for Class 1

impairments (see Table I) are used to prune the candidate

path space (the path space corresponds to column (b) of

Table II). Then the algorithm considers the currentutilization state of the network and uses analytical models

for Class 2 impairments (see Table I) to calculate the

interference among lightpaths. The selection process of

the lightpath is slightly altered. From the set of candidate

lightpaths the algorithm selects the shortest-hop, most

used wavelength lightpath that does not turn infeasible

some of the already established lightpaths. Since the last

criterion can be time consuming we set a limit to the

number of lightpaths that are checked (this limit is set to

5 in the simulation experiments).

C. Multicost IA-RWA algorithm

Assuming that the network supports m wavelengths, the

multicost IA-RWA algorithm presented in [5] uses the

utilization state of the network in order to calculate a cost

vector per link l that has 1+4.m cost parameters,

Vl = (dl, lG , 2

'1',lσ , 2

'0',lσ ,

where lG , 2

'1',lσ , 2

'0',lσ and

lW are vectors of size m that

record the gain, noise variance of bit 1 and bit 0, and the

utilization per wavelength. To calculate the noise

variances of bit 1 and bit 0 for all wavelengths on all

links we use analytical models to account for ASE, XT,

XPM and FWM. These vectors can be calculated offline(in-between connections).

Similarly to the link cost vector, a path has a cost vector

with 1+4.m parameters, in addition to the list of labels of

the links that comprise the path. The cost vector of p can

be calculated by the cost vectors of the links l=1,2,..,n,

that comprise it as follows:

1, 0,( , , , , ,* )p pp pp pd GV W pσ σ= =

, (1, 2,..., ),&

⎛ ⎞⎛ ⎞⎜ ⎟⎜ ⎟⎝ ⎠⎜ ⎟

⎜ ⎟⎛ ⎞⎜ ⎟⎜ ⎟⎝ ⎠⎝ ⎠

∑ ∑ ∑

Using this cost vector a path, we can calculate the Q

factor of the available lightpaths over that path. To do so

we use the noise variance vectors 2

'1',lσ , 2

'0 ',lσ and account

also for eye penalties (due to PMD, SPM/CD and FC).

Eye penalties depend on the path (identified by the last

parameter of the cost vector) and are calculated offline, in

between connections.

The multicost algorithm consists of two phases:

Phase 1: In the first phase, the algorithm computes the set

Pn-d of non-dominated paths from the given source to all

network nodes (including the destination). This algorithm

can be viewed as a generalization of Dijkstra’s algorithm

that only considers scalar link costs. The basic difference

is that instead of a single path, a set of non-dominated

paths between the origin and each node is obtained. Twomechanisms are used to prune the solution space and

reduce the running time of the algorithm. As the paths are

extended by adding new links, we combine the cost

parameters to calculate the Q-factor of candidate

lightpaths and make unavailable those with unacceptable

QoT. We also do not extend paths that have no free

wavelengths. Finally, we use a domination relationship to

prune paths that are worse with respect to all parametersthan other calculated paths to the same end-node.

Phase 2: In the second phase of the algorithm we apply

an optimization function or policy f(Vp) to the cost vector,Vp, of each path p∈Pn-d. The optimization policy f has to

be monotonic in each of the cost components, and yields

a scalar cost per path and wavelength (that is, per

lightpath) in order to select the optimal one. Various

optimization policies that correspond to different IA-

RWA algorithms are presented and evaluated in [5]. For

this study we assume that we use the shortest-hop, most

used wavelength policy, which is also used in the k-SPalgorithms presented above. When making a routing and

wavelength assignment decision, we check if the

establishment of the new lightpath will turn infeasible

some of the already established lightpaths. In case this

happens, there are two options: (i) reroute the connections

that are turned infeasible or (ii) resort to the second

selection choice (the second best lightpath with respect to

the optimization policy), and if this also turns some

established lightpaths infeasible, resort to the third

choice, and so on. To obtain a fair comparison, in this

study we have assumed that we follow the latter approach

and set a limit on the number of candidate lightpaths that

are checked (in particular, in the simulation experiments

the limit was set to 5). Note that in order to evaluate the

effect of the new lightpath on the already established

ones, we use the link cost vectors and the associated

operators described above for a rapid and efficient way to

perform this calculation.

The reason we use a multicost algorithm is threefold.

The first is that we do not use complicated analytical

formulas to account for the interference among lightpaths

during the execution of the algorithm, but we pre-

calculate the noise variances of each wavelength on each

link and keep these values in cost vectors. The cost vector

of each path is then calculated by combining the cost

vectors of the links that comprise it, using simple and

quick operations so that the algorithm runs fast. The

second reason derives from the multicost algorithm’s

nature. The lightpaths calculated have, by the definition

of algorithm operations, acceptable quality of

transmission performance so the IA-RWA problem issolved in a joint manner. Third, having found the

complete set of candidate lightpaths we can explore the

whole lightpath space and apply any optimization policy

when selecting the optimal solution.

In Section VI, we compare the performance, in terms of

connection blocking and execution time, of the three

online IA-RWA algorithms presented. Our results

quantify the benefits of the actual interference approach.

V. WORST-CASE VERSUS ACTUAL-CASE DESIGN OF

TRANSLUCENT NETWORKS IN THE PRESENCE OF PLIS

In this section we focus on translucent networks and onthe number of regenerators required in such networks to

serve a given traffic matrix, under the worst or actual

interference assumptions. In translucent optical networks,

regenerators are employed at some but not all the network

nodes. Some of the connections established are routed

transparently, while others, typically those served by

lengthy paths, may need to utilize one or more

regenerators to restore their signal’s quality. The offline

IA-RWA algorithms proposed for these networks decide

the lightpaths but may also select the regeneration sites

and the number of regenerators that need to be deployed

on these sites, so as to serve the given traffic matrix.In order to provision the network we compare two

different approaches that are based on the same IA-RWA

algorithm presented in [15]. In particular, we compare (i)

a worst case interference IA-RWA algorithm, where the

physical layer constraints are confronted by over-

provisioning the network in terms of regenerators

required, with (ii) an IA-RWA algorithm that calculates

the actual interference among lightpaths, relaxing in this

way the demand for regenerators at the cost of an

increased algorithmic complexity. In both approaches a

traffic matrix is given as the input to the algorithm andthe number of regenerators required to serve this traffic is

recorded as the output of the algorithm.

The IA-RWA algorithm we use under both the worst-

case and the actual-case interference approaches, consists

of three phases. In the first phase, the connection

demands are distinguished into those that can be served

transparently and those that are served using regenerators.

In the actual interference approach, in order to find the

pairs of transparently connected regeneration sites it is

assumed that the network is empty and that only Class 1

impairments affect the QoT of the paths. In contrast, in

the worst case interference approach, it is assumed thatthe network is fully loaded. In both approaches the

quality of transmission evaluation estimator module (Q-

Tool) developed within the DICONET project [16] is

used for assessing the QoT of lightpaths. Next, the non-

transparent connections are transformed into a sequence

of transparent connections by routing them through a

series of regenerators. To do so, the algorithm formulates

a virtual topology problem. The virtual topology consists

of the original network’s regeneration sites, with (virtual)

links between any pair of transparently connected

regeneration sites. Each virtual link of the paths chosen in

the virtual topology to serve a connection, corresponds to

a transparent sub-path (lightpath) in the physical topology

(Figure 3). The algorithms used for routing the non-transparent traffic demands in the virtual topology, are

based on a k-shortest path algorithm, with link costs

defined in two different ways:

1. Virtual-Hop (VH) shortest path algorithm. In

this algorithm all the links of the virtual graph have

cost equal to 1, and the cost of a virtual path is equal to

the number of regenerators it crosses. The optimal

virtual path is the one consisting of the fewest

regenerators (virtual hops).

2. Physical-Hop (PH) shortest path algorithm.

Here the cost of a virtual link is equal to the number of

physical links (physical hops) it consists of. With this

definition, the optimal virtual path is the one thattraverses the minimum number of physical nodes.

Then the algorithm selects the routes to be followed by

non-transparent connections by minimizing one of the

following: i) the maximum number of regenerators used

among all network nodes, or ii) the total number of

regenerators used in the network, or iii) the number of

regeneration sites. To perform this optimization, the

virtual topology problem is formulated as an integer

linear program (ILP).

By the end of the first phase the initial traffic matrix is

transformed into a new traffic matrix whose source-

destination pairs can, in principle, be transparentlyconnected.

In the second phase, when the actual interference

approach is followed, an IA-RWA algorithm for

transparent networks is applied, with input the

transformed transparent traffic matrix, in order find the

RWA solution. On the other hand, in the worst

interference approach, an impairment unaware RWA

algorithm is applied. This is because the fully loaded

network assumption applied in the first phase of the

algorithm, results in all lightpaths having acceptable QoT.

Finally, in the third phase of the algorithm, which isnecessary only for the actual interference approach, the

connections that were rejected in the second phase due to

physical-layer blocking are rerouted through the

remaining (unused in the first phase) regenerators.

In general, the CAPEX (Capital expenditure) and

OPEX (Operational expenditure) of a translucent network

depend not only on the number of wavelengths but also

on the number of regenerator sites and regenerators used.

The basic IA-RWA used [15], distinguishes between

minimizing the maximum number of regenerators used

among the sites and minimizing the total number ofregenerators used or minimizing the number of

regeneration sites. Each of these objectives can be used to

obtain good solutions, depending on the criterion that we

want to optimize. In addition, in our work, the use of the

worst and the actual interference assumptions introduces

a trade-off between the fast execution time and the over-

provisioning of the resources on one hand and the higher

execution time and efficient use of the available resources

on the other. This trade-off is examined in the simulation

results that follow.

VI. SIMULATION EXPERIMENTS

In this section we compare the performance of the worstand the actual interference approaches under the

transparent and translucent network scenarios.

A. Transparent Networks

We compared the performance of three online IA-RWA

algorithms outlined in the previous section: (i) the k-SP

worst-case-interference algorithm with k=5, (ii) the k-SP

actual-interference algorithm with k=5, and (ii) the

multicost algorithm. The topology used in our

simulations was the DTnet topology of Figure 1, with

capacity per wavelength assumed to be 10Gbps. The

physical layer parameters were taken from deliverable

D2.1 of Diconet [16]. We assumed W=16 available

wavelengths per fiber link. We used a random traffic

generator to produce connection requests according to a

Poisson process (rate λ requests/time unit) with

exponentially distributed durations (average 1/µ time

units) and uniformly distributed source-destination nodes.

The network load is defined as λ/µ (in Erlangs). For each

examined load 5000 connections were generated.

Figure 2(a) shows the blocking ratio as a function of

the network load. The multicost algorithm exhibits thebest blocking performance with the performance of the k-

SP actual-interference algorithm coming quite close. The

difference between the multicost and the k-SP actual-

interference algorithm is due to the larger path space that

the multicost algorithm explores. Typically, the multicost

algorithm corresponds to the k-SP actual-interference

algorithm with infinite k, with the path space adjusted and

pruned precisely according to the utilization of thenetwork and the QoT of the calculated lightpaths so as to

have acceptable running time. On the other hand the

difference between the k-SP worst-case-interference and

k-SP actual-interference is more than one order of

magnitude for light loads and decreases as the load

increases. This is expected, since as the network load

increases, the routing options that can be explored by the

k-SP actual-interference algorithm are reduced due to theunavailability of wavelengths. Figure 2(b) shows the

average execution time of the algorithms. As expected the

average execution time of k-SP worst-case-interference

algorithm is the lowest. However, from this graph we can

see that the k-SP actual-interference and the multicost

algorithms also have acceptable execution time that is

kept less than 0.15 sec. This good running time is due to

the sophisticated and quick way that we use to evaluatethe interference among lightpaths and the limit we have

set on the repetition of this process.

B. Translucent Networks

We carried out a number of simulation experiments,

evaluating the performance of several offline IA-RWAalgorithms for translucent networks under both the worst

interference assumption and the actual interference

assumption. The network topology used in our

simulations was the Geant-2 network, shown in Figure 3,

which is a candidate translucent network, as identified by

the DICONET project [16] with 34 nodes and 54bidirectional links (for our simulations we assumed 108

directional links) and a realistic traffic matrix considered

of a total of 400 connections. We assumed W=80

available wavelengths per fiber link. All single-hop

connections were able to be served transparently, but

some multi-hop connections were not, making the use of

regenerators necessary. We assumed that the number of

regeneration sites is not restricted; that is, every node is

capable of accommodating regenerators. It was up to the

proposed algorithms to solve the regeneration placement

problem, in order to decide the regeneration sites and the

number of regenerators to deploy on each site.

Hamb urg

Be rlin

Ha nnover

Br emen

Esse n

Düsseldo rf

Fr ankfur t

Nürnbe rg

St uttgar t

München

Leipzig

Dortmund

Hamb urg

Be rlin

Ha nnover

Br emen

Esse n

Düsseldo rf

Fr ankfur t

Nürnbe rg

St uttgar t

München

Leipzig

Dortmund

Hamb urg

Be rlin

Ha nnover

Br emen

Esse n

Düsseldo rf

Fr ankfur t

Nürnbe rg

St uttgar t

München

Leipzig

Dortmund

Figure 1: DTnet topology used in the simulation experiments.

1.E-04

1.E-03

1.E-02

1.E-01

1.E+00

80 100 120 140 160 180 200

Load (Erlangs)

Blockingratio

k-SP, w orst-case

k-SP, current state

multicost, current state

80 100 120 140 160 180 200Load (Erlangs)

Averageexecutiontim

e(sec)

k-SP, w orst-case

k-SP, current state

multicost, current state

Figure 2: (a) Blocking ratio and (b) average execution time as a function

of network load assuming W=16 available wavelengths

In Figure 4 we graph the total number of regenerators

and the total number of regeneration sites required in thenetwork, to reach zero blocking. The performance of each

algorithm is closely related to the metric it minimizes. In

our results we observe that the PH-based algorithms need

a smaller number of wavelengths to reach zero blocking,

but utilize more regenerators and regeneration sites. On

the other hand, the VH-based algorithms need more

wavelengths to reach zero blocking, but make better use

of the regenerators. In particular, PH algorithms calculate

paths that minimize the number of physical hops utilized,

which tends to give good wavelength utilization

performance, since shorter physical hop paths utilize less

links/wavelengths. This is the reason shorter physical hoppaths are widely used in pure (without impairments)

RWA problems. However, these paths are not directly

related to the virtual topology and thus the ILP algorithm

that runs over the virtual topology does not produce the

best results in relation to its objectives (minimization of

the regenerators). On the other hand, the VH-based

algorithms use as input virtual paths, which are not

related to the physical topology. As a result, the ILP

algorithm, though it places efficiently regenerators in the

network, it selects longer physical hop paths that waste

wavelength resources.

Moreover, as depicted in Figure 4 the IA-RWA

algorithms for translucent networks that use estimates of

the actual interference exhibit better performance when

compared to the algorithms that provision the network

under the worst interference assumption. The worst

interference-based algorithms need to use considerably

more regenerators (more than twice and in many cases

even more) in order to satisfy the same demand matrix.The difference in the required number of regenerators can

be explained as follows. Algorithms under the worst

interference assumption, overuse the available resources

in order to minimize the physical layer blocking, since

they are based on a quite pessimistic assumption that will

only occur if the network is fully loaded. On the other

hand, by using the actual interference approach, network

over-provisioning is relaxed and make better usage of the

network resources. Based on the results of Figure 4, it isbeyond any doubt that using sophisticated algorithms for

network provisioning that account for the interference

among lightpaths is an efficient way to reduce the waste

of resources.

With respect to execution times, for offline traffic that

pertains to the planning phase of the network, there are no

strict time requirements. Since the problem is NP-hard,

acceptable running time usually means that we are able totrack solutions, which is particularly difficult for large

NP-hard problems. In our case, a time limit of a few

hours (5 hours) was set for all offline experiments and the

execution time of the IA-RWA algorithms under the

actual interference assumption were always within limit.

VII. CONCLUSIONS

Due to certain physical effects, routing decisions made

for one lightpath affect and are affected by decisionsmade for other lightpaths. To establish a lightpath for a

new connection we explored two approaches. One

approach is to select a lightpath that has acceptable

quality of transmission (QoT) under the worst case

interference assumption, guaranteeing that the lightpath

will be feasible independently of the establishment of

future connections. This approach is appealing because of

its simplicity and the fact that it does not require any

Figure 3: The non-transparent connection request between source-

destination pair (s, d) can be broken into four transparent sub-path

requests: s-R3, R3-R12, R12-R10 and R10-d. Each of the three sub-path

requests can be served using a different wavelength.

229269 256

660611

PH-ILPsum PH-ILPmax PH-ILPsites PH-ILPsum-

worst case

PH-ILPmax-

worst case

PH-ILPsites-

worst case

214 222 226

631602

VH-ILPsum VH-ILPmax VH-ILPsites VH-ILPsum-

worst case

VH-ILPmax-

worst case

VH-ILPsites-

worst case

Figure 4: Total number of regenerators and total number of regeneration

sites for 400 connection demands, W=80 available wavelengths andunrestricted regeneration sites. (a) PH and (b) VH based algorithms.

checks on the effect the establishment of a new

connection will have on existing connections; however, it

tends to overuse the wavelength and regenerator

resources. The second approach is to take into accountthe current network utilization and perform a cross layer

optimization between the network and physical layers.

The second approach explores a larger path space and

performs significantly better, in terms of blocking ratio

and resources utilized, than algorithms that follow the

worst case interference approach, but has increased

complexity. We proposed and evaluated sophisticated

techniques that follow the second cross-layeroptimization approach. Our results indicated that we can

keep the execution times low, comparable to those of

algorithms that follow the worst-case assumption, and

also obtain significant performance benefits.

ACKNOWLEDGMENT

This work has been supported by the European

Commission through DICONET project [16].

REFERENCES

[1] H. Zang, J. P. Jue, B. Mukherjee, “A Review of Routingand Wavelength Assignment Approaches for Wavelength-

Routed Optical WDM Networks”, Optical NetworksMagazine, Vol. 1, 2000.

[2] Y. Huang, J. Heritage, B. Mukherjee, “Connectionprovisioning with transmission impairment considerationin optical WDM networks with highspeed channels”, J. of

Lightwave Technology, Vol. 23, No. 3, 2005.

[3] J. He, M. Brandt-Pearce, Y. Pointurier, S. Subramaniam,

"QoT-Aware Routing in Impairment-Constrained OpticalNetworks", pp. 26-30, IEEE Globecom, 2007.

[4] V. Anagnostopoulos, C. Politi, C. Matrakidis, A. Stavdas,

“Physical layer impairment aware routing algorithms basedon analytically calculated constraints”, Optics

Communications, 2006.[5] K. Christodoulopoulos, K. Manousakis, E. A. Varvarigos,

M. Angelou, I. Tomkos, "A Multicost Approach to OnlineImpairment- Aware RWA", IEEE International Conferenceon Communications (ICC), 2009.

[6] G. Shen, W. D. Grover, T. H. Cheng, S. K. Bose, “Sparseplacement of electronic switching nodes for low-blocking

in translucent optical networks”, J. of Optical Networking1, pp. 424-441, 2002.

[7] Y. Ye, T. Y. Chai, T. H. Cheng, C. Lu, “Novel algorithm

for upgrading of translucent optical networks”, OpticsExpress, Vol. 11, No. 23, pp. 3022-3033, 2003.

[8] X. Yang, B. Ramamurthy, “Dynamic Routing inTranslucent WDM Optical Networks: The Intra-Domain

Case”, J. of Lightwave Technology, Vol. 23, No. 3, 2005.[9] Y. Ye, T. Y. Chai, C. Lu, “Routing and Wavelength

assignment algorithms for translucent optical networks”,

Optics communications, pp. 233-239, 2004.[10] S. Pachnicke, T. Paschenda, P. Krummrich, “Assessment

of a constraint-based routing algorithm for translucent 10Gbits/s DWDM networks considering fiber nonlinearities”,J. of Optical Networking, Vol. 7, No. 4, pp. 365-377, 2008.

[11] Y. Ye, T. Y. Chai, T. H. Cheng, C. Lu, “Algorithms for thedesign of WDM translucent optical networks”, Optics

Express, Vol. 11, No. 22, pp. 2917-2926 , 2003.[12] G. Shen, W. D. Grover, T. H. Cheng, S. K. Bose, “Sparse

placement of electronic switching nodes for low blocking

in translucent optical networks”, OSA Journal of Optical

Networking, Vol. 1, pp. 424-441, 2002.[13] X. Yang, B. Ramamurthy, “Sparse Regeneration in

Translucent Wavelength-Routed Optical Networks:Architecture, Network Design and Wavelength Routing”,

Photonic Network Communications, Vol. 10, No. 1, 2005.

[14] M. A. Ezzahdi, S. A. Zahr, M. Koubaa, N. Puech, M.Gagnaire “LERP: a Quality of Transmission Dependent

Heuristic for Routing and Wavelength Assignment inHybrid WDM Networks”, International Conference on

Computer Communications and Networks (ICCCN), 2006.[15] K. Manousakis, K. Christodoulopoulos, E. Kamitsas, I.

Tomkos, E. A. Varvarigos, " Offline Impairment-Aware

Routing and Wavelength Assignment Algorithms inTranslucent WDM Optical Networks", J. of Lightwave

Technology, Vol. 27, No. 12, pp. 1866-1877, 2009.[16] Dynamic Impairment Constraint Network for Transparent

Mesh Optical Networks (DICONET), www.diconet.eu

Konstantinos Christodoulopoulos received a Diploma in Electrical

and Computer Engineering from the National Technical University ofAthens, Greece, in 2002, the M.Sc. degree in Advanced Computing

from Imperial College London, U.K., in 2004, and the Ph.D. degree

from the Computer Engineering and Informatics Department, Universityof Patras, Greece, in 2009. His research interests are in the areas of

protocols and algorithms for optical networks and grid computing.

Panagiotis Kokkinos received a Diploma in Computer Engineering andInformatics in 2003 and an M.Sc. degree in Integrated Software and

Hardware Systems in 2006, both from the University of Patras, Greece.

He is currently a Ph.D. student at the Department of ComputerEngineering and Informatics of the University of Patras. His research

activities are in the areas of ad-hoc networks and grid computing.

Konstantinos Manousakis received the Diploma degree from the

Computer Engineering and Informatics Department, University of

Patras, Greece, in 2004 and the M.Sc. degree in Computer Science and

Engineering from the Computer Engineering and InformaticsDepartment in 2007. He is currently a Ph.D. candidate at the same

department. His research activities focus on optimization algorithms for

high speed and optical networks.

Emmanouel (Manos) Varvarigos received a Diploma in Electrical and

Computer Engineering from the National Technical University ofAthens in 1988, and the M.Sc. and Ph.D. degrees in Electrical

Engineering and Computer Science from the Massachusetts Institute of

Technology in 1990 and 1992, respectively. He has held faculty

positions at the University of California, Santa Barbara (1992-1998, as

an Assistant and later an Associate Professor) and Delft University of

Technology, the Netherlands (1998-2000, as an Associate Professor). In

2000 he became a Professor at the department of Computer Engineering

and Informatics at the University of Patras, Greece, where he heads the

Communication Networks Lab. He is also the Director of the Network

Technologies Sector (NTS) at the Research Academic Computer

Technology Institute (RA-CTI), which through its involvement in

pioneering research and development projects, has a major role in the

development of network technologies and telematic services in Greece.

His research activities are in the areas of high-speed network, protocolsand architectures, distributed computation and grid computing.

ICBR-Diff: an Impairment Constraint Based Routing Strategy with Quality of Signal

Differentiation

Amornrat Jirattigalachote, Paolo Monti, Lena Wosinska Royal Institute of Technology KTH, School of Information and Communication Technology, Isafjordsgatan 22,

Electrum 229, 164 40 Kista, Sweden Email: {amornrat, pmonti, wosinska}@kth.se

Kostas Katrinis and Anna Tzanakaki

Athens Information Technology (AIT), 19.5 Markopoulo Av., P.O. Box 68, 19002 Peania, Greece Email: {kkat, atza}@ait.edu.gr

Abstract—Latest advances in Wavelength Division Multiplexing (WDM) technology make it possible to build all-optical transparent networks, which are considered to be able to satisfy the rapidly growing capacity demand. However, in a transparent WDM network the optical signal transmitted from a source to a destination node might be degraded due to physical layer impairments induced by transmission through optical fibers and components. Several Impairment Constraint Based Routing (ICBR) algorithms have been proposed to consider physical layer impairments during the connection-provisioning phase in order to prevent selecting a lightpath with poor signal quality. However, these algorithms support only a single quality of transmission threshold for all connection requests, while next generation networks and the future Internet are expected to support a variety of services with potentially disparate QoS requirements. In this paper, we propose the use of bit error rate (BER) as a differentiation of service parameter for connection requests in optical WDM networks. This is achieved through the use of ICBR, whereby various BER thresholds are set depending on the QoS requirements for accepting/blocking the connection requests during the connection-provisioning phase. The evaluation results reveal that significant network performance improvement in terms of connection blocking can be achieved, compared to non-differentiated conventional routing and wavelength assignment (RWA) and ICBR algorithms. Index Terms—Impairment Constraint Based Routing (ICBR), differentiation of services, signal quality, physical layer impairments, connection provisioning, transparent optical networks.

I. INTRODUCTION

Transparent WDM networks constitute a promising solution to cater for the rapid growing of bandwidth

demand in next generation networks and the future Internet. In such networks, the signal is transported from source to destination in the optical domain via all-optical channels (also referred to as lightpaths [1]) with capacity that can reach 100 Gbit/s [2]. These lightpaths do not require any intermediate optoelectronic processing, thus reducing the number of costly O/E/O converters. In addition, all-optical networks offer bit rate, signal format, and protocol transparency.

Many routing and wavelength assignment (RWA) algorithms [3][4] have been proposed in order to assign a lightpath, i.e. a route and wavelength, to a connection request. One objective of these RWA algorithms is to reserve the minimum amount of network resources for all connection requests. By minimizing the amount of reserved resources, the reduction in connection blocking, i.e. the blocking of a connection request due to the lack of available resources in the network, is expected.

However, these RWA algorithms base their routing decisions only on the availability of network resources, and assume that optical fibers and components are ideal. But in practice, an optical signal might be degraded by physical layer impairments inherent in the fiber segments and optical components, in the absence of intermediate optoelectronic conversions offering signal regeneration [5]. This effect clearly introduces the need to consider physical layer impairments during the connection-provisioning phase, i.e. the need to solve the so-called Impairment Aware Routing and Wavelength Assignment (IA-RWA) problem. The objective of solving the IA-RWA problem is not only to minimize the amount of reserved network resources, i.e. to reduce the connection blocking, but also to guarantee the required quality of transmission level, e.g. measured in terms of bit error rate (BER), for each connection request.

Physical layer impairments can be divided into linear and non-linear [6][7]. Linear impairments do not depend on the signal power and affect each channel individually. A list of the most important linear impairments includes: Amplified Spontaneous Emission (ASE) noise, Group

Parts of this paper appeared in Proc. of IEEE ICTON/RONEXT’09; Manuscript received January 29, 2010; accepted March 22, 2010.

Velocity Dispersion (GVD) and Polarization Mode Dispersion (PMD). ASE is due to optical amplification, whereby optical amplifiers degrade the optical signal to noise ratio (OSNR) of the transmitted optical signal. GVD causes a pulse broadening effect and is the result of different spectral components traveling at different velocities. PMD causes optical power variation and signal distortion due to the fiber birefringence where each of the two orthogonally polarized modes travels slower than the other. Non-linear impairments on the other hand, are triggered due to the response of the optical fiber to intense optical signals [8][9]. Specific nonlinear effects that are important in the context of optical signal transmission in fibers are Self Phase Modulation (SPM) and Cross Phase Modulation (XPM) caused by the dependence of the fiber refractive index on the intensity of the applied electromagnetic field. SPM is an effect through which an optical signal causes degradation of itself in the form of spectral broadening and pulse chirping. On the other hand, XPM introduces spectral broadening of the optical signal through its interaction with the other channels also transmitted in the same fiber. Another important nonlinear effect is Four Wave Mixing (FWM), by reason of which unwanted optical components are generated due to propagation of the optical channels through the fiber. Nonlinear effects are in general emphasized for high signal power levels and longer transmission distances. In summary, both linear and non-linear impairments are highly dependent on the fiber characteristics which in turn are sensitive to length, temperature and age.

Several Impairment Constraint Based Routing (ICBR) algorithms [10][11] have been proposed in the literature to solve the IA-RWA problem. The work in [10] presents two different ICBR algorithms. One of them first selects the lightpath with the shortest physical distance and then it checks the lightpath signal quality. The second one picks the first available wavelength on the selected route, and then performs a signal quality check on that specific wavelength. If the signal quality of the selected wavelength does not satisfy the signal quality threshold, the next available wavelength is chosen. Both ICBR algorithms in [10] compare the signal quality of the selected lightpath against a single quality threshold and consider only linear physical layer impairments, namely PMD and ASE noise. The work in [11] considers, in addition to linear physical layer impairments, also the effect of non-linear impairments, i.e. SPM, XPM, and FWM. The authors present an ICBR algorithm that characterizes and orders the wavelengths based on their quality factor (Q-factor) values and it chooses the lightpath with the highest Q-factor. This lightpath quality is then compared against a single quality threshold.

The family of ICBR algorithms outlined above are usually referred to as Impairment-Aware Best-Path (IABP) routing algorithms because of the way they make routing decisions, i.e., they always assign each connection request the least impaired lightpath, regardless of the signal quality requirement of the connection request. However, a variety of services, such as peer-to-

peer (P2P) applications, high definition television (HDTV) and Audio Video On Demand (AVOD), require different quality of transmission, bandwidth and delay, while these ICBR algorithms treat all connection requests in the same manner. In other words, they support only a single signal quality threshold, uniformly to all connection requests (i.e. the lowest BER). As a consequence, these single-threshold approaches may unnecessary block connection requests that could sustain signal degradation of some extent, i.e. higher BER than the single threshold imposed by these flat-service ICBR schemes. In addition, for those connection requests that are accepted, such single-threshold approaches may overprovision network resources, i.e. connection requests with low BER threshold are unnecessarily assigned well performing fibers, with a potentially detrimental effect on the overall connection blocking.

To overcome these deficiencies, this paper proposes and evaluates a novel ICBR algorithm, referred to as ICBR-Diff, supporting differentiation of services at the BER level. In the proposed algorithm, various BER thresholds are considered for accepting/blocking connection requests in the connection-provisioning phase, depending on the signal quality requirements of the connection requests. This allows for a better use of the available resources by offering optical signal quality that is good enough for a specific connection request. Simulation results on the Pan-European test network topology and NSFNet topology, show significant improvement in terms of connection blocking, compared with conventional Shortest Path and IABP algorithms.

The rest of this paper is organized as follows. Section II presents the transmission link model and physical layer impairments model used for proposed algorithm. In Section III, the proposed ICBR-Diff algorithm is described. The performance evaluation of the ICBR-Diff algorithm is presented in Section IV. Finally, Section V provides some concluding remarks.

II. ESTIMATING THE TRANSMISSION QUALITY OF THE CONNECTION REQUESTS

This section describes the assumptions made and provides details on (i) the approach and (ii) the equations used to estimate the transmission quality of a connection request routed along a given path in the network.

It is assumed that the WDM network has an arbitrary physical topology (mesh). The WDM mesh is modeled as a graph G=(N,L) where N represents the set of network nodes, i.e. OXCs, and L represent the set of network links. It is assumed that every network link consists of a set of bidirectional fibers, F, where each fiber carries a set of wavelengths, W.

Given a connection request r between a source and a destination node, the objective of any impairment constrained based routing strategy is to find a path p ∈ G=(N,L) able to accommodate some predefined quality of transmission requirements. In this work the quality of transmission is measured in terms of bit error rate (BER). The BER computation may vary depending on the assumptions made in modeling a network link and on the

optical impairments that are considered. A detailed description of the assumptions used to model each network link is provided next.

A. Network link model Amplification and dispersion management [12][13]

techniques are some of the available approaches used to increase the reach of optical transparent transmissions, and to guarantee an acceptable level of signal quality at the receiver (i.e. the destination node). In this work, these techniques, combined with the approach presented by G. Markidis et al. [11], are used to model the transmission link connecting two nodes in the network, as shown in Fig. 1.

Figure 1. Network link architecture.

Fig. 1 shows that the assumed network link consists of

a sequence of single mode fiber (SMF) spans. Their number may vary according to the physical distance separating the two nodes. Erbium-doped fiber amplifiers (EDFAs) are inserted following each fiber span in order to compensate for the power loss induced by the fiber. For long-haul DWDM transmission careful dispersion management is critical. In this context, a non-resonant dispersion map that utilizes pre-compensation and post-compensation aiming to achieve optimum system performance is used. Accumulated dispersion is also compensated by allocating in line dispersion compensating fiber (DCF). Based on this model, the Q-factor of each link is computed. This computation is then be used in the estimation of the transmission quality on an entire path p, as explained in the next sections.

B. Estimating the transmission quality of a link in the network

In this work, the effect of physical layer impairments is quantified using the quality factor Q, defined as [11]. The Q-penalty factor includes both linear and non-linear physical layer impairments, namely ASE noise, the combined effects of SPM/GVD and optical filtering, FWM, and XPM. ASE, FWM and XPM are calculated by assuming that they follow a Gaussian distribution. For the combined SPM/GVD and optical filtering effects, they are quantified through an eye closure penalty metric. The Q-penalty factor on the k-th link can be expressed as:

kXPMkFWMkASEkpenalty pen

,σσσ ++

= , (1)

where is the relative eye closure attributed to

SPM/GVD and optical filtering effects, is the

electrical variance of ASE noise, while and

are the electrical variance of FWM and XPM

induced degradation respectively. Details on each one of these terms are provided next.

keyepen ,

2,kASEσ

2,FWMσ k

2XPMσ

1) ASE noise: the electrical variance of the ASE noise [7] of the k-th link, calculated at the end of the cascade of the optical amplifiers, can be expressed as:

eASEavgkASE B

BPPR ⋅= 22

, 4σ , (2)

where is the responsivity of the receiver, is the

average signal power at the receiver, is the ASE power of EDFAs [7], and are the optical and the electrical bandwidth of the receiver, respectively.

R avgP

2) SPM/GVD: both SPM and GVD are responsible for the broadening of the pulse. For this reason their effect can be combined and considered together[14][15]. By convoluting the transfer function of the fiber with the assumed frequency chirping at the transmitter, the eye closure penalty ( keyepen , ) due to SPM/GVD at the end of the link can be calculated as:

peneye,k = Achirpped (ω)∗H fiber (ω), (3) where is the transmitted pulse with frequency chirping (modeled as in [6][7][15]), and

is the transfer function of the fiber.

)(ωchirppedA

)ω(fiberH

3) FWM: the electrical variance of FWM of the k-th link can be obtained by using the following expression [16]:

⎟⎟⎠

⎞⎜⎜⎝

⎛++= ∑ ∑∑

II IIIaacabm

IabctkFWM PPPPR

812 22

,σ , (4)

where is the responsivity of the receiver, is the transmitted signal power of each channel, the first summation term represents the FWM power [17] in the case that all channels including the considered channel (m) are using different frequencies (

cf mba fff ≠≠≠ ), the second summation represents the FWM power when

ca ff bf ≠≠ but mc ff = , and the last summation represents the FWM power when mfcb ffaf ≠≠= .

4) XPM: The electrical variance of XPM of the k-th link can be expressed by [18],

σXPM ,k,a2 = P2 1

2π(x) ⋅ (y) ⋅ (z)[

−∞

∫b=1,b≠a

∑ dω , (5)

2, )()( ωIMabXPMHx = ,

. )()( ωfilteroptHy = , )()( ωbPSDz =

where P is the average channel power, is the total number of SMF spans, is the XPM-induced intensity modulation frequency response originated by the pump channel b on the probe channel a,

N)(, ωIM

abXPMH

)(. ωfilteroptH is the transfer function of the optical filter at the receiver, and )(ωbPSD is the power spectral density of pump

It is assumed that the connection requests in R are grouped in different classes, {C1,C2,…,Cv}, where each class corresponds to a different BER threshold according to their signal quality requirements. Also, the assignment of connection class Ci occurs in a one-to-one basis. Let

{ }jim

jijiji pppP ,,2

, ,...,,= represent the set of m candidate routes between source node i and destination node j, and

{ }jim

ji pBER ,,2

,...,ji

ppP BERBERBER ,

represent the BER

(computed as in Equation (7)) of each of the candidate routes in P , .

channel b. The detailed calculation of can be found in [19].

)(, ωIMabXPMH

C. Estimating the transmission quality of a path in the network

Once a connection request is routed along a given path, the resulting transmission quality needs to be assessed. A path p ∈ G=(N,L) is assumed to be a sequence of network link, where the Q-penalty factor of each link is expressed as in Equation (1). The Q-factor of a path p can be expressed as:

Now, assume that irPp ∈ is a candidate route for

routing ri. With any conventional ICBR approach, the BER of p (denoted as BER ) is compared against a single and predefined BER threshold (denoted as ). This threshold, usually very stringent (e.g. 10-15, 10-16), is chosen a priory and is used to check the feasibility of all ri ∈ R. During this process, no attention is paid on what is the actual difference, in terms of BER, between and

thrsBER

BER thrs BER . As such, this approach may potentially lead to over-provisioning of network resources. The ICBR-Diff algorithm, on the other hand, tries to avoid over-provisioning of resources by trying to find a route p that (1) satisfies the signal quality

requirement of ri (denoted as ) and, at the same time, (2) matches as closely as possible the value of

, i.e.,

∑ ∑ ∑

∈k pkkXPM

2,σ++

∈ ∈pk pkFWMkASE

tpeyep

σσQ , (6)

peneye,pwhere is the transmitted signal power, tP is the

eye closure penalty due to SPM/GVD effects (as defined in Equation (3)) calculated basing on the total length of the path, is the total electrical variance of

ASE noise (as defined in Equation (2)) of the path, and are the total noise variance

of FWM and XPM of the path (as defined in Equation (4) and (5)), respectively.

σASE ,k2

k∈p∑

,kk∈p∑σ

k∈p∑ FWM

2 σ XPM ,k2

The BER value of a connection request routed along p can be derived from Equation (6) as follows:

⎥⎦⎤

⎢⎣⎡ −=

irdirsjirdirs

BERBERp )(),()(),(

min ⎟⎟

⎞⎜⎜⎝

erfcBER . (7) . ⎥⎦

⎤⎢⎣⎡ >⎟

⎠⎞⎜

⎝⎛ −∧ 0)(),(

iirdirs

rp BERBER

This in turn leads to a more efficient use of network resources (and consequently a lower blocking probability) as shown in the example that follows.

III. ICBR-DIFF: AN IMPAIRMENT CONSTRAINT BASED ROUTING ALGORITHM WITH SERVICE DIFFERENTIATION

In this section, the general idea of ICBR-Diff algorithm is presented, followed by a detailed description of the algorithm.

Fig. 2 shows an example where the ICBR-Diff algorithm and a conventional IABP algorithm are compared. Let R={r1, r2, r3} be the set of connection requests from node A to node B. Let also 10-15, 10-9, and 10-15 be their respective signal quality requirements.

Let R={r1,r2,…,rq} represent the set of connection requests that needs to be provisioned in G=(N,L). The source and destination nodes of ri ∈ R are denoted by s(ri) and by d(ri) respectively.

Figure 2. Path selection example.

We also assumed a specific assignment of routing

requests to classes of service, namely: {r1, r3} ∈ C1 with BER≤10-15 and {r2} ∈ C2 with BER≤10-9.

Let { }BABABABABABA pppppP ,5

, ,,,,=

1510 −=thrsBER1CBER

9102 −=

be the set of m=5 candidate routes between node A and node B, with one wavelength available per route (i.e. W=1). The set of BER values of each candidate route, i.e. , is shown in Fig. 2. As previously explained, the IABP algorithm fixes the BER threshold, BERthrs, to a constant value (in the example ), while the ICBR-Diff uses the threshold or

, depending on the class, which the connection request being processed belongs to.

BAPBER,

1510 −=CBER

When r1 arrives, IABP assigns route to r1, i.e.

the least impaired route that satisfies . ICBR-Diff, on the other hand, assigns which is the route

with the highest BER value in that satisfies the signal quality requirement of the connection request of

C1. When r2 arrives, IABP chooses while ICBR-

Diff selects . Notice that r2 belongs to C2 and

requires a BER value less than 10-9, while provides a BER value equal to 10-16. Finally, when r3 arrives, IABP is not able to find a feasible route, since none of routes left with resources in

BAp ,1

thrsBER

BAp ,2

pBAp ,

P , satisfies the signal quality imposed by . Instead, ICBR-Diff is able to assign to r3. The rationale from the example is the following: ICBR-Diff avoids unnecessary blocking

thrsBERB,Ap1

Initialize network topology information and physical parameters; foreach incoming connection request ri ∈ R of class Ci T(N,L’) = current topology status; foreach link k ∈ L’ cost link k = Qpenalty,k; end ( ))(),()(),(

2)(),(

1)(),( ,...,, iiiiiiii rdrs

mrdrsrdrsrdrs pppP =Path_Finding_Module(T,m);

if =)(),( ii rdrsP Ø then Block ri; break; end p = Path_Selection_Module(

by selecting routes that closely match the connection request quality of service requirement.

The pseudo code of the ICBR-Diff algorithm is presented in Fig. 3. The algorithm starts with an initialization phase where the network topology information is collected, e.g. number of nodes, number of links, link lengths, link capacities, and all physical parameters required for the calculation of the Q-penalty factor of each link. Graph T(N,L’), with L’⊆ L, represents the current resource usage in the network, i.e. a link belongs to L’ if and only if it has resources available. Initially, graph T(N,L’) is equal to graph G(N,L). The algorithm assigns the Q-penalty factor, Qpenalty,k, calculated as in Equation (1) as the cost for each link. Graph T(N,L’) is then weighted by using these link cost values. After assigning link costs, up to m alternative routes for each connection request are computed by running the Dijkstra algorithm on the weighted graph T(N,L’). Fig. 4 presents the pseudo code of the algorithm used to find m alternative routes. If there is a route p with at least one common wavelength available on every link, this route is stored in the set of candidate routes

)(),( ii rdrsP . The cost of links of p is then doubled, the weight of the links in graph T is updated with the new link costs, and additional candidate routes are computed. Otherwise, if no route is found or if there is not a common available wavelength on p, the connection request is blocked.

Next, the BER of each candidate route in set )(),( ii rdrsP is calculated using Equation (7). Fig. 5 shows the pseudo code of the path selection process implemented in the ICBR-Diff algorithm. With ICBR-Diff, different signal quality thresholds are considered according to the signal quality requirements of the incoming connection requests.

)(),( ii rdrsP ,Ci);

if p = NULL then Block ri; else Update T; return p ; end end

Figure 3. Pseudo code of ICBR-Diff.

When starting the path selection process (Fig. 5), the signal quality threshold is set to be the specific value based on the class the connection request belongs to. The BER of each candidate route in set )(),( ii rdrs

jp )(),( ii rdrsP is compared against the signal quality threshold of the class of the considered connection request (i.e. ). The candidate route with the highest BER that satisfies the signal quality requirement of the connection request is selected. Finally, the first wavelength in the list of available wavelengths of the selected route is chosen to route the lightpath for the corresponding connection request.

iCBER)(),( ii rdrs

Path_Finding_Module(T,m){ h=1; for h≤m p = Dijkstra algorithm(T); if p = Ø then break; else if common available wavelength on p then ph = p; foreach link ∈ p new link cost = link cost * 2; end else ph = Ø; end end; h = h + 1; end return {p1,p2,…,pm}; }

Figure 4. Pseudo code of path finding module.

Path_Selection_Module( )(),( ii rdrsP ,Ci){ Compute BER P s ( ri ),d ( ri = {BER

p1s ( ri ),d ( ri ) ,BER

p 2s ( ri ),d ( ri ) ,..., BER

p ms ( ri ),d ( ri ) }

p = NULL; MIN = ∞; foreach

)(),(irdirs

irdirsj

Pp BERBER ∈

if and iirdirs

Cp BERBER <)(),( MINBERBER i

irdirsj

Cp <−)(),( then

)(),( ii rdrsjpp = ;

MIN = iirdirs

Cp BERBER −)(),( ;

end end return p ; }

Figure 5. Pseudo code of path selection module.

IV. PERFORMANCE EVALUATION

In this section, we first introduce the definitions and assumptions used in the evaluation of ICBR-Diff. Then some simulation results are presented.

A. Definitions and assumptions It is assumed that the bandwidth demand of each

connection request is one wavelength unit and that wavelength conversion capability is not available, i.e., wavelength continuity constraint is enforced while solving the routing problem. Furthermore, our experiment model assumes random and dynamic incoming connection requests that are sequentially served without prior knowledge of future incoming connection requests.

TABLE I.

SYSTEM PARAMETERS USED FOR IMPAIRMENTS CALCULATION

Parameters Values Central wavelength 1553.6 nm (193.1 THz) Channel spacing 50 GHz Channel bit rate 10 Gbit/s Spontaneous-emission factor of in-line EDFA (nsp)

Spontaneous-emission factor of EDFA before entering node (nsp)

Attenuation of SMF 0.25 dB/km Attenuation of DCF 0.5 dB/km Nonlinear index coefficient of SMF

2.6x10-20 m2/W

Nonlinear index coefficient of DCF

3.5x10-20 m2/W

GVD parameters of SMF 17x10-6 s/m2

GVD parameter of DCF -80x10-6 s/m2 Dispersion slope of SMF 0.085x103 s/m Dispersion slope of DCF -0.3x10-3 s/m Effective area of SMF 65x10-12 m2 Effective area of DCF 22x10-12 m2 Optical bandwidth (Bo) 40 GHz Electrical bandwidth (Be) 7 GHz Maximum length of SMF 80 km The incoming connection requests follow a Poisson

distribution, while source/destination pairs are randomly chosen with equal probability (uniform distribution) among all network nodes. Connection holding time is exponentially distributed with the mean equal to 6 time units.

As mentioned earlier, ICBR-Diff supports differentiation of services, whereby connection requests are divided into two distinct classes with regard to their signal quality requirements, i.e. Class-1 connection requests that require higher signal quality in terms of maximum tolerated BER, and Class-2 connection request that can tolerate higher signal degradation than Class-1. Throughout our simulations, Class-1 connection requests require BER less than 10-15 and Class-2 connection requests require BER less than 10-9. Furthermore, two configurations of traffic mix for Class-1 and Class-2 connection request are considered: (i) 30% of the overall traffic being of Class-1, while 70% of the traffic being of Class-2 and (ii) equal share of Class-1 and Class-2, i.e. each class amounting to 50% of the overall traffic.

These signal quality requirements of Class-1 and Class-2 connection requests were chosen based on available IP traffic measurement [20][21]. It is shown that streaming media traffic, i.e. traffic requiring higher signal quality (Class-1), accounts for much less of the total bandwidth utilization, whereas peer-to-peer (P2P) and World Wide Web (WWW) traffic, i.e. traffic requiring lower signal quality (Class-2), is dominating.

The parameter values used for the calculation of physical layer impairments in this work are listed in Table I.

B. Simulation results To evaluate our ICBR-Diff algorithm, we used (i) the

Pan-European test network topology (COST 239) [22], which comprises 11 nodes and 26 bidirectional fiber links with 16 wavelengths per fiber, and (ii) the NSF network

(NSFNet) consisting of 16 nodes and 24 bidirectional fiber links with 16 wavelengths per fiber (Fig. 6).

For benchmarking purpose, we also evaluate two other provisioning algorithms, namely shortest path and IABP. In the shortest path algorithm, the physical link distance is used as the link cost. The candidate route with the shortest physical distance in the set P is selected first, then the BER of the selected route is calculated against a single signal quality threshold ( ). With IABP, each link is assigned a cost equal to Qpenalty,k, and the BER of every candidate route in the set P is calculated and compared against a single signal quality threshold,

. The candidate route with the lowest BER value that satisfies the single signal quality threshold is selected. For both shortest path and IABP, the value of

is set to be equal to 10-15. These two approaches do not support service differentiation and thus connection requests are blocked if there is no route with BER less than , irrespectively of the class that the connection request belongs to. In the case of ICBR-Diff algorithm, Class-1 connection request is blocked if there is no available lightpath connecting source and destination which exhibits BER less than 10-15, whereas Class-2 connection request is blocked if there is no lightpath with BER less than 10-9.

thrsBER

BER thrs

The total blocking probability shown in Fig. 7 accounts for both blocking due to insufficient resources, i.e. no wavelength is available, and due to the impairment constraints, when the candidate routes cannot meet the signal quality requirement. Additionally, the probability of connection blocking due to insufficient resources and the probability of connection blocking due to impairments are separately showed in Fig. 8 and Fig. 9, respectively. In both COST 239 and NSFNet topologies, the results show significant improvement in terms of connection blocking achieved by the ICBR-Diff algorithm, compared to shortest path and IABP routing. When Class-1 and Class-2 connection requests account for the 30% and 70% of total requests respectively, the benefit obtained by our ICBR-Diff algorithm in COST 239 topology is almost up to an order of magnitude in terms of total connection blocking compared to shortest path routing, and up to 81% compared to IABP algorithm, While the benefit achieved by ICBR-Diff in NSFNet is up to 61%, compared to both shortest path and IABP approaches, since the connection blocking in the case of IABP is almost identical with the connection blocking in the case of shortest path algorithm. In the second case considering a traffic mix of Class-1 and Class-2 connection requests being equally weighted, the benefit reduces to 89% and 67% in the case of COST 239, and to 45% in the case of NSFNet.

By comparing the connection blocking in the case of shortest path routing against the IABP, we can see that the connection blocking reduction achieved by IABP algorithm is significant in the COST 239 topology, while the connection blocking of shortest path routing and IABP algorithms are almost identical in the NSFNet

Figure 6. Pan-European COST 239 test network (left), and NSF network (right).

Figure 7. Total blocking probability versus load.

Figure 8. Insufficient-resource blocking versus load.

topology. This is because the diversity of the set of candidate routes is higher in COST 239 than in NSFNet, due to the higher nodal degree (see Table II and Fig. 6). Thus, the benefit, in terms of connection blocking, achieved by IABP in COST 239 is larger than that in NSFNet. Furthermore, the link length in COST 239 is relatively short, which in turns means the less impaired link. Thus, in IABP, there might be the possibility that a candidate route which is not the shortest one, in terms of physical length, is able to satisfy the signal quality requirement of the connection request. By considering all candidate routes rather than choosing the shortest one first, i.e. the shortest distance, IABP already gives the performance improvement in terms of connection blocking compared to shortest path routing. However, in ICBR-Diff, a higher improvement in connection blocking can be achieved by both considering all candidate routes and providing only the sufficient-signal quality lightpaths to the connection requests.

Fig. 7 shows that in COST 239 a slight performance improvement is gained by our ICBR-Diff, i.e. ICBRDiff- 100% Class-1 in the figure, compared to IABP algorithm. To further improve the connection blocking, differentiation among connection requests is applied, and the simulation results (Fig. 7) show that significant performance improvement in terms of connection blocking can be achieved using the proposed ICBR-Diff algorithm.

In the case of NSFNet where the diversity of the set of the candidate routes is low due to the low nodal degree and the link length is relatively long compared to COST 239, shortest path routing and IABP end up selecting the same route. Because of the long link length, the route that satisfies the signal quality threshold is the shortest route also. By providing only the good-enough lightpath without the consideration of the differentiation of connection requests, i.e. ICBRDiff-100% Class-1, ICBR-Diff gives similar performance compared to IABP. This is also because of the low diversity of the set of the candidate routes and the long link length in NSFNet. However, when considering also the differentiation of

Figure 9. Impairments blocking versus load.

TABLE II.

COMPARISION BETWEEN COST 239 AND NSFNET

Topology COST 239 NSFNet Number of nodes 11 16

Number of bidirectional links 26 24 Average node degree 4.73 3

signal quality requirements of connection requests during the connection provisioning phase, the considerably connection blocking improvement can be obtained by our ICBR-Diff algorithm.

The comparison between the simulation results of COST 239 and NSFNet shows that the ICBR-Diff algorithm is able to facilitate improved utilization of the network resources in both high and low nodal degree topologies, i.e. COST 239 and NSFNet respectively, compared to shortest path routing and the IABP approach. Moreover, in the low nodal degree and long link length topology, i.e. NSFNet, our ICBR-Diff algorithm can still significantly improve the performance in terms of connection blocking, while IABP cannot.

V. CONCLUSIONS

In this paper, we proposed a novel Impairment Constraint Based Routing (ICBR) algorithm with differentiation of services based on the BER of a lightpath for each connection request. In contrast to the existing ICBR and Impairment-Aware Best-Path (IABP) algorithms, in our approach the signal quality requirement of the connection request, in terms of maximum tolerated BER, is considered as a routing constraint during the connection-provisioning phase. Simulation results indicate significant improvement in connection blocking in both COST 239 and NSFNet compared to shortest path routing and the IABP approach. By assigning lightpaths to the connection requests, having acceptable BER performance with respect to the predefined thresholds, and by avoiding to choose lighpaths with a lower BER value compared to the threshold required, our ICBR-Diff algorithm is able to offer more efficient resource utilization. This is due to

that the best performing lighpaths remain available for future use when new and more demanding connection requests, in terms of BER, are to be set up. Furthermore, another interesting finding is that our ICBR-Diff has a high flexibility to different network topologies.

The network performance improvement offered by the proposed ICBR-Diff has been obtained by avoiding connection blocking due to unnecessary high signal quality constraint.

Simulation results show that in the topology where the node degree is low, i.e. the diversity of the set of the candidate routes is low, and the link length is long (NSFNet), our ICBR-Diff is still able to give significant performance improvement in terms of connection blocking, while IABP gives similar performance to shortest path routing.

ACKNOWLEDGMENT

This work was supported by the Network of Excellence “Building the Future Optical Network in Europe” (BONE), funded by the European Commission through the 7th ICT-Framework Programme and the EUREKA/CELTIC project “Management Platform for Next Generation Optical Networks” (MANGO).

REFERENCES [1] I. Chlamtac, A. Ganz, and G. Karmi, “Lightpath

Communications: a novel approach to high bandwidth optical WANs”, IEEE Transactions on Communications, vol. 40, no. 7, pp. 1171-1182, 1992.

[2] J. Berthold, A.A.M. Saleh, L. Blair, and J.M. Simmons, “Optical Networking: past present and future”, IEEE/OSA Journal of Lightwave Technology, vol. 26, no. 9, pp. 1104-1118, 2008.

[3] R. Ramasawami and K.N. Sivarajan, “Routing and Wavelength Assignment in All-Optical Networks”, IEEE/ACM Transaction on Networking, vol. 3, no. 5, pp. 489-500, Oct. 1995.

[4] D. Cavendish, A. Kolarov, and B. Sengupta, “Routing and Wavelength Assignment in WDM Mesh Networks”, in Proc. Conf. Globecom 2004, Dec. 2004.

[5] J. Strand, A.L. Chiu, and R. Tkach, “Issues For Routing In The Optical Layer”, IEEE Communications Magazine, vol. 39, no. 2, pp. 81-87, Feb. 2001.

[6] R. Ramasawami, and K.N. Sivarajan, “Optical networks: a practical perspective”, 2nd edition, Morgan Kaufmann, San Francisco, 2002.

[7] G.P. Agrawal, “Fiber-Optic Communication Systems”, 3rd edition, A John Wiley & Sons, INC., Publication, New York, 2002.

[8] M. Farahmand, D. Awduche, S. Tibuleac, and D. Atlas, “Characterization and representation of impairments for routing and path control in all-optical networks”, in Proc. of National Fiber Optic Engineers Conference (NFOEC), Dallas, TX, Sep. 2002.

[9] B. Ramamurthy, D. Datta, H. Feng, J.P. Heritage, and B. Mukherjee, “Impact of transmission impairments on the teletraffic performance of wavelength-routed optical networks”, IEEE/OSA Journal of Lightwave Technology, vol. 17, no. 10, pp. 1713-1723, Oct. 1999.

[10] Y. Huang, J.P. Heritage, and B. Mukherjee, “Connection Provisioning With Transmission Impairment Consideration in Optical WDM Networks With High-

Speed Channels”, IEEE/OSA Journal of Lightwave Technology, vol. 23, no. 3, pp. 982-993, Mar. 2005.

[11] G. Markidis, S. Sygletos, A. Tzanakaki, and I. Tomkos, “Impairment Aware based Routing and Wavelength Assignment in Transparent Long Hual Networks”, in Proc. Conf. on Optical Network Design and Modeling (ONDM), May 2007.

[12] D. Penninckx, G. Charlet, J.-C. Antona, and L. Noirie, “Simple engineering rules for a transparent waveband-based optical backbone network”, OSA Journal of Optical Networking, vol. 2, no. 2, pp. 38-45, Feb. 2003.

[13] I. Tomkos, M. Vasilyev, J.-K. Rhee, A. Kobyakov, M. Ajgaonkar, and M. Sharma, “Dispersion map design for 10 Gb/s ultra-long haul DWDM transparent optical networks”, in Proc. OECC’02, Yokohama, Japan, Jul. 2002.

[14] M.I. Hayee, and A.E. Willner, “Pre- and Post-Compensation of Dispersion and Nonlinearities in 10-Gb/s WDM Systems”, IEEE Photonics Technology Letters, vol. 9, no. 9, pp. 1271-1273, Sep. 1997.

[15] M. Stern, J.P. Heritage, R.N. Thurston, and S. Tu, “Self-Phase Modulation and Dispersion in High Data Rate Fiber-Optic Transmission Systems”, IEEE/OSA Journal of Lightwave Technology, vol. 8, no. 7, pp. 1009-1016, Jul. 1990.

[16] K. Inoue, K. Nakanishi, K. Oda, and H. Toba, ”Crosstalk and Power Penalty Due to Fiber Four-Wave Mixing in Multichannel Transmissions”, Journal of Lightwave Technology, vol. 12, no. 8, pp. 1423-1439, Aug. 1994.

[17] K. Inoue, “Four-Wave Mixing in an Optical Fiber in the Zero-Dispersion Wavelength Region”, IEEE/OSA Journal of Lightwave Technology, vol. 10, no. 11, pp. 1553-1561, Nov. 1992.

[18] S. Pachnicke, S. Spalter, J. Reichert, and E. Voges, “Analytical Assessment of the Q-factor due to Cross-Phase Modulation (XPM) in Multispan WDM Transmission Systems”, in Proc. SPIE, vol. 5247, pp. 61-70, Sep. 2003.

[19] A.V.T. Cartaxo, “Cross-Phase Modulation in Intensity Modulation-Direct Detection WDM Systems with Multiple Optical Amplifiers and Dispersion Compensators”, IEEE/OSA Journal of Lightwave Technology, vol. 17, no. 2, pp. 178-190, Feb. 1999.

[20] Traffic Measurements and Models in Multi-Service Networks project, Celtic project, “TRAMMS IP Traffic report no. 1, April 2008”, http://projects.celtic-initiative.org/tramms/.

[21] Traffic Measurements and Models in Multi-Service Networks project, Celtic project, “TRAMMS IP Traffic report no. 3, June 2008”, http://projects.celtic-initiative.org/tramms/.

[22] P. Batchelor et al., “Study on the implementation of optical transparent transport networks in the European environment-Results of the research project COST 239”, Photonic Network Communications, vol. 2, no. 1, pp. 15-32, 2000.

Amornrat Jirattigalachote obtained her Bachelor’s degree in Telecommunication Engineering from King Mongkut’s Institute of Technology Ladkrabang, Bangkok, Thailand in 2004, and her Master degree in Electrical Engineering (specialized in Photonics) from the Royal Institute of Technology (KTH), Stockholm, Sweden in 2008.

Kostas M. Katrinis earned the diploma in Computer Engineering from the Computer Engineering and Informatics Dept., University of Patras, Greece in 2000 and the Ph.D. in Technical Sciences from the Swiss Federal Institute of Technology (ETH Zurich), Zurich, Switzerland in 2006.

She worked as a Tooling Software Development Engineer at Western Digital (Thailand) Co., Ltd. from 2004 to 2005. In 2005, she started working as a System Engineer with Aeronautical Radio of Thailand Ltd. From 2008, she is a Ph.D. student at KTH in the area of Optical Networking. Her main research interests presently are on Physical Layer Impairment Aware Routing in Multi-domain Multi-granularity Optical Networks and Power Efficiency in Survivable WDM Networks.

Paolo Monti received a Laurea degree in Electrical Engineering (2001) from the Politecnico di Torino, Italy, and a Ph.D. in Electrical Engineering (2005) from the University of Texas at Dallas (UTD).

From 2006 to 2008 he worked as a Research Associate of the Open Networking Advance Research (OpNeAR) Lab at UTD. He joined the

Royal Institute of Technology (KTH) in September 2008 where he is currently an Assistant Professor in the School of Information and Communication Technology (ICT/FMI) and a member of the Next Generation Optical Networks (NEGONET) group. He co-authored more than thirty papers published in international journals and presented in leading international conferences. His research interests include network planning, protocol design, performance evaluation and optimization techniques for both optical and wireless networks.

He is currently with the Network Design and Services (NDS) group at the Athens Information Technology (AIT), Athens, Greece. In the past, he has worked as a research assistant at the Communication Systems Group at ETH Zurich in the area of multimedia communication. He has authored or co-authored numerous publications in international journals and conferences and two book chapters. His research interests are in the area of network planning, sustainable networking, trustworthy communication and network performance evaluation.

Dr. Katrinis is a member of the Technical Chamber of Greece since 2000.

Lena Wosinska received her Ph.D. degree in photonics and the Docent degree in optical networking from the Royal Institute of Technology (KTH), Stockholm, Sweden, in 1999 and 2008, respectively.

She joined KTH in 1986, where she is currently an Associate Professor with the School of Information and

Communication Technology (ICT), heading a research group in optical networking (Next Generation Optical Networks NEGONET) and coordinating a number of national and international scientific projects. Her research interests include optical network management, reliability and survivability of optical networks, photonics in switching, and fiber access networks.

Anna Tzanakaki is an Associate Professor at the Athens Information Technology, where she is leading the Network Design and Services research group. She is also an adjunct faculty member of Carnegie Mellon University, USA. She has obtained a BSc degree from the University of Crete, Greece, an MSc and a PhD both from the University

of Essex, UK. She was a co-founder and a senior engineer of ilotron ltd, a spin-off from the University of Essex, involved in the design of systems for WDM optical networks. Following ilotron, she joined Altamar Networks, a subsidiary of Ditech Communications, as a principal engineer responsible for optical architecture and system design. She is a co-author of over 100 publications in international journals and conferences. She is a co-inventor of 1 granted and 11 published patents. She is a senior member of the IEEE and several Technical Program Committees. Her research interests include optical wavelength, burst and packet switched networks, cross-layer network design and traffic provisioning as well as network convergence in support of telecommunications and IT services.

Dr. Wosinska has been involved in a number of professional activities, including Guest Editorship of the following special issues that appeared in the OSA Journal of Optical Networking: High Availability in Optical Networks; Photonics in Switching; Reliability Issues in Optical Networks; and Optical Networks for the Future Internet. Since 2007, she has been an Associate Editor of the OSA Journal of Optical Networking, and since April 2009 she serves in the Editorial Board of the IEEE/OSA Journal of Optical Communications and Networking. Since 2005, she has been a General Chair of the Workshop on Reliability Issues in Next Generation Optical Networks (RONEXT), which is a part of the IEEE International Conference on Transparent Optical Networks (ICTON). She serves on the technical program committees of many international conferences.

Feedback Based Load Balancing, Deflection Routing and Admission Control in OBS

Networks

Sébastien Rumley, Christian Gaumier Laboratoire de Télécommunications (TCOM), Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland

Oscar Pedrola, Josep Solé Pareta

CCABA, Universitat Politècnica de Catalunya, Spain

Abstract—The Optical Burst Switching (OBS) paradigm allows statistical multiplexing directly at the optical layer. Thus, OBS networks are suited to carry traffic demands varying in either the short or long term. Due to the lack of buffering, burst contention due to short term variations can only be mitigated through deflection routing. For longer term variations, higher order mechanisms such as dynamic flow-balancing or flow shaping are generally proposed. In this paper, a unified scheme, based on a feedback mechanism combined with deflection routing and admission control is introduced to handle all types of traffic variations. The use of only one single scheme simplifies the architecture of OBS networks and enhances its flexibility. The validity of our technique is supported by simulation results. Index Terms—Optical burst switching, deflection routing, feedback mechanism, load balancing, admission control.

I. INTRODUCTION

Optical networks already carry the vast majority of the long-distance communications. Their role is nevertheless promised to be even greater with the advent of ultra-high bandwidth home accesses. However, while conceived in the past to carry constant or at least stationary traffic, they have nowadays to deal with an increasing part of traffic demands varying both in the short and long term. For instance, in digital TV long term variation are due to the users switching on or off the TV whereas short term variations are caused by compression algorithms. Future optical networks are required to cope with these significant traffic variations without loss of quality.

Optical Burst Switched (OBS) networks, given their ability to achieve statistical multiplexing directly at the optical layer, are considered as a promising approach to efficiently satisfy future communication demands. Unlike Optical Circuit Switched Networks (OCS), they offer sub-wavelength bandwidth granularity. They remain, however, simpler than their Optical Packet Switched (OPS) counterpart, whose implementation is still questionable [1].

In the basic OBS scheme, traffic is aggregated in large packets called bursts. Before sending a burst in the network, an associated control packet is sent in advance. The control packet reserves the resources required to

dispatch the burst toward destination. Emitted on a dedicated channel of reduced throughput, control packets are easy to decode, unlike OPS headers. Sent in advance, they let enough time to core nodes to prepare for the burst arrival. In this way, bursts travel transparently in the network, at very high rates [2].

This conventional OBS scheme presents nevertheless several major drawbacks. Due to its bufferless nature, it cannot solve transient congestions by shortly delaying some of the contending burst, as in a classical packet switching paradigm. This weakness causes unavoidable burst losses as soon as traffic is not determinist. Furthermore, the pre-allocation of resources leads to low channel utilisation ratios.

Various approaches have been investigated and combined to both reduce as much as possible the burst loss rate and to maximise the throughput. They can be distinguished in two classes. Within the first class, local node resources only are used to solve contentions. These resources can be optical buffers [3] or complex schedulers [4]. Local approaches are beyond the scope of this article, and will not be further discussed.

Within the second class, approaches include mechanisms which encompass the whole network. Two of these methods – Deflection Routing (DR) and Load-Balancing (LB) – rely on the frequent existence of several routes between a pair of nodes. Thus, if one route is saturated, traffic can be deflected over an alternate path. However, DR and LB differ in the scale at which they operate. Specifically, DR reroutes on a per burst and per hop basis (one individual burst is rerouted over one single hop) whereas LB achieves a per flow and per path rerouting (all the bursts of one flow are rerouted over a whole path).

The per burst aspect of the DR permits to individually deflect bursts to an alternate route, which is worthy when casual bursts fail to obtain a reservation on a given channel, due to short term variations [5]. Unfortunately, when facing long term traffic variations, which cause bursts to be systematically contended at one output port, DR might simply transpose the congestion to another link rather than solve it [6]. In this latter case, the traffic should instead be routed differently in the network, to avoid systematic congestions. This is exactly what Load-

Balancing (LB) achieves, by limiting the number of certain bursts injected into each particular route. LB can be done statically or dynamically. In the static case, traffic is balanced according to an extrapolation based forecast. In the dynamical case, flows are repetitively estimated on the fly, and the load-balancing is achieved according to these successive estimations.

The problem with LB is that too tight route restrictions may favour transient contentions. Hence, if a burst cannot be forwarded to the next hop of its predefined route, it must be dropped. It results that LB and DR have to be combined to mitigate both short term and long term congestions. Unfortunately, however, their simultaneous usage may lead to problematic situations: whilst BL tries to restrict the routes to particular ones, DR attempts to find alternative paths. Synchronizing both mechanisms might be particularly difficult, since they often differ in the way that they are implemented. While deflection mechanisms are intrinsically related to core nodes, load-balancing ones are more likely located at edge nodes, and sometime even operate at a different layer.

To avoid the previously discussed problems while keeping the benefits of both mechanisms, a unified technique called Adaptive Deflection Routing (ADR) is proposed in this article. This scheme, operating at core nodes, routes the traffic dynamically, selecting the output port which ensures, for the rest of the path, the best quality for current traffic conditions. If the preferred output port is congested, deflection is performed towards the second most efficient output port, and so forth.

Moreover, the transmission quality can be further improved by limiting the access of the bursts which will or are likely to be dropped anyway. This mechanism is referred to as Admission Control (AC) or congestion avoidance. The ADR scheme presented in this paper also integrates AC capabilities. Thus, a burst might be dropped rather than forwarded, even if the local port is not congested, if the risk of latter contention is high.

ADR estimates the risk and quality associated to each forwarding operation using feedback messages exchanged between core nodes. When a node voluntarily drops a burst or fails to reserve a channel, a negative feedback is sent to all core nodes previously visited by the burst. Similarly, positive feedbacks are sent when a burst reaches its final destination. In this way, a core node receiving a feedback knows a posteriori the consequences of one of its past choices.

In Section 2, the principles on which the ADR scheme relies are discussed with respect to other approaches proposed in the literature. Implementation details are given in Section 3. The behaviour of the scheme is analysed through simulations in Section 4. Section 5 provides some conclusions.

II. MODELLING OF ADR

ADR consists of four components: Deflection Routing, Load-Balancing, Admission Control and Feedback Based Adaptation. Each of them has already been largely studied in the past. The present section reviews them briefly.

A. Deflection Routing in OBS

Deflection Routing can be superposed to any other routing scheme. If a burst can be forwarded to the output port (or one of the output ports) defined by the primary routing scheme, no deflection occurs. If on the contrary all the primary ports are busy, the burst is deflected to an alternative port.

The selection of the alternative port can be achieved in various ways. In the most simple deflection scheme, an alternative port is selected randomly or according to an arbitrary order among the idle ones. In more complex cases, a list of alternative ports is assigned to each primary port or, better, to each burst final destination.

Lists permit either fixing an order for alternatives, either limiting the number of potential deflection ports, or both. The criterion for ordering and limitation is generally the distance separating the next node of the output port from the destination. Hence, deflections implicitly lengthening the burst journey are both less likely to be selected (due to ordering) and more likely to be excluded (by limitations).

The exclusion or selection of an output port highly depends on the burst remaining offset time trem (or, equivalently, on the remaining offset time unit urem=trem/tp, where tp is the burst header processing time). Indeed, each deflection operation is equivalent to a switching operation and consumes one unit. Thus, a burst deflected too many times might be dropped owing to an insufficient trem (insufficient offset time problem [20]). Avoiding deflecting bursts on routes incompatible with the remaining offset will lead to better performances. Lists may thus be setup for each potential final destination and possible value of urem. Besides, additional units of offset times can be granted to bursts at emission. These bonus units allow a burst to be deflected more times, and may help a particular burst to find its way in the network. However, this will increase the resources consumed by each burst, which can be counterproductive [6]

It is worth pointing out that the use of ordered lists prioritizing particular deflection ports implicitly affects the way traffic flows in the network. Thus, a type of local load balancing can be achieved by this means. This load-balancing can even be achieved dynamically if core nodes exchange feedbacks with their neighbours [7].

In our ADR scheme, deflection is achieved according to lists recomputed after each feedback reception. However, contrarily to [7], these feedbacks are not issued by neighbouring nodes only, but by all nodes traversed by a traffic flow.

B. Routing and Load-Balancing in OBS

Considerable work has been achieved to study how to optimally balance burst flows in a network. A survey of the basic routing schemes is available in [8].

The literature on LB can be separated in two groups, depending on the question: is an a priori knowledge of the traffic pattern assumed? If the answer is yes, load-balancing can be achieved using optimisation techniques. Several models have been developed to evaluate the global performance of an OBS network, depending on its

topology, traffic matrix, and core node architecture. These models can be derived to perform linear [9] or non-linear optimisation [10]. In many situations, however, a priori knowledge of the traffic pattern is unavailable. This is generally true when traffic fluctuates too much (any estimate becomes either obsolete very quickly, or is too averaged to represent the real traffic). These situations call for dynamic flow balancing (DFB).

DFB in OBS can be achieved in a similar way that traffic engineering is performed in (G)MPLS networks. Information of the congestion in the network is given to a client of the OBS layer, by means of a flooding protocol dispatching link state information messages (e.g. OSPF). Based on this information, the client layer selects itself the routes of its bursts, avoiding congested links [11]. This solution is, however, problematic when link state information is obsolete. Therefore, network state information must be available prior to proceed to any further change (in particular, after a previous rerouting operation).

Rather than letting the client layer manage the routing, this last can be operated by the OBS layer itself. This approach permits taking into account OBS specific congestion metrics. It also allows a decentralisation of the routing decisions. Hence, individual edge nodes [12]-[15] or core nodes [7],[16]-[19] can decide independently on which route (for edge nodes) or on which output port (for core nodes) bursts must be sent.

Edge-decision based schemes generally take into account the whole network status, allowing optimal routing. However, as in the MPLS scheme, the time required to collect network state may be problematic. Core-decision based schemes, on the contrary, base their decisions on the state of a limited part of the network (e.g. their immediate neighbourhood). This provides shorter reaction times after traffic changes, since up-to-date information is available more rapidly. However, a local view is highly likely to lead to suboptimal configurations [13].

In both cases, individual nodes may react independently but simultaneously, which will produce oscillations in the congestion [12]. More generally, when load-balancing is applied, a trade-off must inevitably be found between very reactive decisions causing oscillations and less reactive ones, keeping the network in a suboptimal state for a longer time.

In our ADR approach, routing decisions, similarly to deflection, are taken by core nodes according to received feedbacks. However, these feedbacks are originated in all other nodes, conferring then a global nature to the ADR routing scheme.

C. Feedback mechanism

Relevant information exchanged between nodes can be referred to as per burst or per link.

In the per burst approach, core nodes simply send feedback each time a burst is either dropped, received, or switched [13]-[15],[18]. The duty of analysing the feedbacks and deducing the congestion state is left to the receiving node. In the per link perspective, each core

node estimates the state of each of its own output links. It then broadcasts this information [7],[12],[19],[20].

The per burst approach permits reporting critical situations such as repeated burst losses, almost immediately. On the contrary, the per link type techniques average the link state over time. Sudden changes may thus need longer time to be detected. On the other hand, the per burst approach is likely to generate more control overhead to dispatch correctly all the feedbacks. Nevertheless, since a signalling channel is required in OBS anyway, the impact of this additional overhead is expected to be moderate [13].

The ADR feedback mechanism is based on a per burst paradigm. The feedback only consists of either ACK or NACK messages, associated with a burst identifier. However, contrarily to other per burst approaches, ADR feedback is sent to all previously visited nodes, and not only to the burst emitting one.

D. Admission Control

While in connection oriented networks, rejected traffic has no impact on the accepted one, this is not the case in datagram-based networks. Prior to be dropped, packets may consume and thus waste network resources. This may dramatically reduce the network total throughput [6]. Mimicking the connection oriented networks, mechanisms voluntarily dropping a part of the incoming datagram at network entrance have been proposed. This protection technique is usually referred to as Admission Control (AC).

AC in OBS can be achieved at either edge or core nodes. The dichotomy between per path, when AC is achieved at edges, and per hop, at cores, appears hence again. AC-equipped edge nodes shape the traffic they inject in the core network. This shaping operation can be performed either by buffering the traffic exceeding a given rate [21], or by dropping it. The targeted maximal output rate can be fixed by static planning. It may also be estimated dynamically, according to feedbacks received from the network, or simply according to the flow history.

There is no utility to perform per hop AC with fixed-routing schemes: all the exceeding traffic on an output port could be, in this case, shaped by the edge node. However, AC at core nodes becomes interesting when DR policies are applied. In fact, DR performs implicitly an admission control if the list of deflection alternatives does not contain all the output ports. Hence, several schemes have been proposed to limit the deflection alternatives according to some network state information [18][22].

The AC mechanism integrated in ADR proceeds similarly. Based on the feedbacks, core nodes include or exclude forwarding options. All forwarding options might even be excluded. In such a case, a burst is said blocked. This differs from the other approaches, which exclude options independently of the network state, and which never exclude them all.

III. ADR IMPLEMENTATION

ADR consists of a routing logic, providing ordered forwarding options whenever a control packet arrives at a core node. Its only restriction concerns the structure of the burst control packet (BCP). BCP must contain, at least, a unique identifier (ID), an indication of the burst remaining offset time, and a history field permitting to record the nodes visited so far. If an emulated OBS [23] architecture is used, the remaining offset time is replaced by a Time-To-Live field. Additionally, ADR requires an access to the signalling channel in order to transmit the feedbacks.

ADR performs four distinct operations: A) record in a local memory its last decisions B) send feedbacks to other nodes C) collect feedbacks to analyse its previous choices D) decide about burst forwarding

A. Decision record

When a core node admits a burst (i.e. schedule a resource reservation), it associates the burst ID with an information triplet (D, R, N) that consists of the burst final destination (D), the burst remaining offset (R), and the selected forwarding port (N). This ID=(D, R, N) association is stored in a local memory.

B. Emission of feedback packets

Each time a burst reaches its destination or is dropped, a feedback is sent to all its previously visited nodes. This feedback follows the path formerly followed by the burst, but in the opposite direction. Feedback packets contain the ID of the corresponding burst, and either an ACK or a NACK flag.

C. Reception of feedback packets

Upon reception of a feedback message, core nodes retrieve the (D, R, N) triplet associated to the feedback ID. Using these three values, it accesses a specific memory structure, called Time Sliding Feedback Counters (TSFCs, described later on), and increments the number of received feedbacks.

TSFCs store all the feedbacks received in the recent past. They are split in Q cells. Each cell stores the positive and negative feedbacks received in a finite amount of time [t, t+∆]. Thus, TSFCs record feedbacks collected during a time Q∆.

The feedbacks received within the time interval [n∆, (n+1)∆] are stored at the address n (mod Q). Periodically, at each tk = k∆, k=Q,Q+1,Q+2…, records taken during time interval [(k-Q)∆, (k-Q+1)∆] are erased, allowing more recent feedback to be stored.

C. Forwarding and dropping decisions

Upon reception of a BCP, a core node calls its ADR routing logic, which immediately refers to its TSFCs. Let us assume that:

• the burst announced by the BCP is destined to d • it has r units of offset remaining • it arrives from node s • current node has M output ports p1,…,pm

then, the TSFCs corresponding to the triplet (d, r, l) are retrieved, with l=1 ... m, l ≠ s, s being deduced from the burst history field (i.e. the TSFCs of all ports except the one the burst comes from). The values

νl= total feedbacks πl=(positive feedbacks/ νl )

are extracted from each retrieved TSFC. If no feedback has been collected yet, πl is set to 1.

In the next step, the unfavourable forwarding choices are excluded. An option j is considered unfavourable if the following conditions are met:

• πj < θπ • νj > θv

where θπ and θv are fixed thresholds. The first condition excludes ports which led to bad results in the past (i.e. those showing a low π ratio). Setting a minimal number of feedback θv avoids excluding a possibility which has not been fully assessed yet.

Once the remaining forwarding alternatives sorted according to the π values, ADR operates similarly to the classical DR scheme. A reservation is attempted on the first item of the list. If this attempt fails, the following items are considered until a reservation is scheduled. If two or more options share the same π value, they are chosen randomly.

If a reservation is achieved on port l, the decision is recorded together with the d and r values, as previously discussed. On the contrary, if all the proposed output links are saturated, or if all ports have been excluded, the BCP is not forwarded, and a negative feedback is sent back. The block diagram of Fig. 1 summarises the ADR operation. This procedure is also exemplified through the following example. Assuming the network topology depicted in Fig. 2, a burst destined to node d=3 with r=2 units of remaining offset arrives at node 1 from node 4. Node 1 has three forwarding options: 0, 2 or 5. Option 5 is excluded because:

• π5=0 < θπ = 0.15 • v5=20 ≥ θv = 30

Feedback arrival

Burst admittedon port l

Burst droppedor atdestination

Retrieve final destination d, remainingoffset r and identifier ID from BCP

Retrieve the corresponding TSFCs,compute π and ν

Exclusion of non-acceptableports, sorting of the acceptable one

Successive trial of reservation on thesorted ports

BCParrival Send feedback to

the node(s) traversed by the

current burst

Forward BCP to l

Add burst corresponding

triplet (d,r,l) in the memory

Triplet memory

Recover the triplet (d,r,l)associated with the feedback identifier

Update the feedback counterscorresponding to d and r

Feedback arrival

Burst admittedon port l

Burst droppedor atdestination

Retrieve final destination d, remainingoffset r and identifier ID from BCP

Retrieve the corresponding TSFCs,compute π and ν

Exclusion of non-acceptableports, sorting of the acceptable one

Successive trial of reservation on thesorted ports

BCParrival Send feedback to

the node(s) traversed by the

current burst

Forward BCP to l

Add burst corresponding

triplet (d,r,l) in the memory

Triplet memory

Recover the triplet (d,r,l)associated with the feedback identifier

Update the feedback counterscorresponding to d and r

Figure 1. Operational block diagram of the ADR scheme.

TCAC = 0.15FCAC = 20

TSFCs at node 1 for d=3 and r=2:- π0= 0.979 (+;- feedbacks : 521;11)- π2 = 0.944 (+;- feedbacks : 120;7)- π5 = 0.0 (+;- feedbacks : 0;20)

Figure 2. Node 1 considers the forwarding options for a burst destined

(a) (b)

Figure 3. (a) SIMPLE and (b) EON topologies. Node 1 tries first to schedule the burst on the link towards 0 (π0 = 0.979). If all the wavelengths of the link are occupied, same operation is made with port towards node 2. If this last attempt fails again, the burst is dropped. In this precise case, no positive feedbacks have been received for next-hop node 5. This is obvious, since route 1-5-0-3 counts 3 hops. No burst holding an offset r = 2 can thus join node 3 following this route.

ADR hence operates solely by sending, counting, and analysing feedbacks. At the initialization stage, only the parameters θπ , θv, C and ∆ have to be set. Moreover, ADR performs all operations (routing, deflection, admission control) in an integrated manner.

IV. NUMERICAL RESULTS

Several simulations have been performed to estimate the performance of our ADR scheme using the JAVOBS [24] simulation tool. For all links, 16 wavelengths at 10 Gbit/s have been assumed. Emission rate is normalized to the link capacity. For an offered rate ρ=1, each node emits a total of 160 Gbit/s, uniformly distributed on the remaining nodes. Traffic is generated according to the Poisson model. Mean burst size is fixed to 1.2 Mbit.

Since the performance of ADR is independent of the scheduling algorithm used, switching time ts and processing time tp are neglected. However, the number of offset unit remaining urem is still set as it would be with a non null processing time.

Simulations have been performed on two different topologies. First, the SIMPLE network (Fig. 3a), which has 6 nodes and 8 links, has been used to analyse in deep the ADR behaviour. In this case, links are assumed to have very short lengths, and thus, no delay is taken into account for feedbacks.

0.25 1 2 3 4 5 7 9 110

offered load(a)

ADR, σ=0

ADR, σ=2

ADR, σ=4

DR, σ=0

DR, σ=2

DR, σ=4

0.25 1 2 3 4 5 7 9 11

offered load(b)

Figure 4: Performance comparison with other routing approaches, for high loads

Second, ADR has been tested in more practical

situations using the EON topology (Fig. 3b [26]), which consists of 28 nodes and 41 links. Real link lengths, corresponding to geographical distances, have been used to simulate transmission delays, for both burst and feedbacks.

A. Simulations on SIMPLE topology

ADR performance is first compared to the well-known Shortest-Path (SP) Routing and to the Deflection Routing (DR). It is assumed that DR systematically tries all forwarding options which, given the remaining offset, permit to reach the destination. The options requiring fewer hops are considered in priority (random selection for equalities).

A parameter σ taking values 0, 2 or 4 represents an offset time supplement granted to each burst. It can be exploited by both ADR and DR. The thresholds θπ = 0.5 and θv = 10 are used with TSFCs of Q=2000 cells and ∆=40µs.

Performances in terms of burst loss ratio (BLR) and carried load are represented in Fig. 4(a) and 4(b). ADR performs better than the other approaches as the load increases. Contrarily to DR, ADR performances do not fall below the SP. Even more, for ρ>5, while the carried load remains stable as the offered traffic increases (SP, DR with σ = 0) or even drops (DR with σ > 0), the ADR carried traffic still increases and gets stable only for extreme loads (ρ>11).

Fig. 5 focuses on the burst loss ratio for lower loads (ρ<2). For almost null rates, SP and DR exhibit low burst ratios. For ρ=0.1, using SP or DR and σ = 0, and for

ρ<0.7 using DR and σ > 0, no burst loss has even been recorded within the simulated time of 200ms. On the contrary, burst loss ratio of ADR is never null, due to two facts. Firstly, ADR has no infinite memory and has to relearn periodically the good forwarding options. Secondly, a minor part of the traffic needs to be lost during the learning process, when forward success probabilities are not yet known. However, as load increases, the percentage represented by this lost traffic decreases. For ρ>1.5, ADR with σ > 0 outperforms again the other schemes.

To apprehend the differences between DR and ADR, the load ρp offered on the port p has been measured, for each port. An average per port offered load

)...(1

1 mPρρρ ++=

has then been computed, P being the total amount of ports and ρi the load of port i. The recorded ρ values are

represented in Fig. 6. Assuming a SP routing and lossless conditions, the

ideal average per port offered load SPρ% can be computed as

( 1)SP N P

ρρ α= ⋅ ⋅−

where ρ/(N-1) is the load carried by each route, P is the

total number of ports in the network and α is the total amount of hops for all routes in the network. Thus,

SPρ% corresponds to the load offered on each route,

multiplied by the number ports (resources) employed by these routes, and averaged over all ports. On the SIMPLE topology, α=46, P=16, and N=6, thus SPρ% = 0.575 for ρ=1

and SPρ% = 2.3 for ρ=4. These values are represented on

Fig. 6 with thin vertical dotted lines. For ρ=1, SPρ is just below SPρ% since SPρ is decreased

by the burst losses [27]. On the contrary, DRρ and ADRρ are

greater thanSPρ% , especially when σ > 0. This is because in

DR and ADR the amount of intermediate nodes (hops) is higher due to the deflection. The offset time supplement σ, contributes obviously to the increase in the number of traversed hops since it extends burst lifetimes.

For ρ=4, SPρ is drastically reduced by the losses and

appears clearly belowρ~

. However, the largest difference

consists in the explosion of DRρ . Indeed, a more loaded

network conducts DR to exhaust all forwarding possibilities before dropping a burst. Giving a high offset time bonus strengthens this effect. This increased amount of extra visited ports explains why performance drops at high loads. With ADR on the contrary, ADRρ is far

below SPρ% , and even belowSPρ . Hence, due to admission

control mechanisms, ADR drops in excess bursts prior to offering them to any port. This action spares capacity for other bursts and explains why performances are not affected.

Fig.7 represents the amount of bursts dropped after a journey of n hops. For ρ = 1, DR and ADR do not differ much in terms of hops traversed by dropped bursts. For

ρ = 4, lost bursts travel a much longer distance before being dropped (for σ = 4, about 1000 burst travelled over 6 hops). On the contrary, ADR blocks earlier these “resource waster”. Since SP does not use offset larger than required and since the longest path includes three hops, no bursts are dropped after 3 hops.

0.1 0.5 1 1.5 2

offered load

ADR, σ=0

ADR, σ=2

ADR, σ=4

DR, σ=0

DR, σ=2

DR, σ=4

Figure 5. Performance comparison with other routing approaches, for low loads.

0 1 20.575 2.3 3 4 5

Average load offered per port

ρ=4 ρ=1

Figure 6: Average per port offered load ρ for SIMPLE topology.

Dotted lines represent ρ~ estimates.

0 1 2 3 4 5 6 7

distance travelled, ρ=1

ADR, σ=0

ADR, σ=2

ADR, σ=4

DR, σ=0

DR, σ=2

DR, σ=4

0 1 2 3 4 5 6 7

distance travelled, ρ=4

Figure 7. Record of the distances (in hops) travelled by dropped bursts.

TABLE I BREAKDOWN OF THE REASONS FOR BURSTS LOSSES

Hops achieved when dropped

σ Drop reason 0 1 2 3 4 5 6 7

No availabilities 42.97% 98.67% 95.41% 38.37% 6.06% 0.00% 46.73%

Blocked 57.03% 1.33% 4.59% 6.40% 0.00% 0.00% 53.24%

Offset exhausted 0.00% 0.00% 0.00% 55.23% 93.94% 100.00% 0.03%

No availabilities 45.50% 99.01% 92.36% 87.29% 91.30% 31.82% 25.00% 0.00%

Blocked 54.50% 0.99% 7.64% 12.71% 8.70% 9.09% 0.00% 0.00%

Offset exhausted 0.00% 0.00% 0.00% 0.00% 0.00% 59.09% 75.00% 100.00%

0 0.1 0.2 0.3 0.4 0.5 0.60.3

θπ (a)

σ=0, θν=0

σ=0, θν=10

σ=0, θν=50

σ=2, θν=0

σ=2, θν=10

σ=2, θν=50

0 0.1 0.2 0.3 0.4 0.5 0.60.45

θπ (b)

Figure 8. Burst loss vs. θπ, for ρ=2 (a) and ρ=3

1 3 5 7 9 11

offered load

θπ=0.5

θπ=0.6

θπ=0.7

θπ=0.8

θπ=0.9

θπ=0.95

Figure 9. Performance of ADR for various θπ values.

Finally, the reasons generating burst drops are

analysed in Table 1. Among the bursts dropped after 0 hops (i.e. at the network entrance), about half is blocked. A blocked burst is not offered to any port prior to be dropped. After one hop, only about 1% of the bursts are blocked. Thus ADR drops exceeding traffic at network entrance, which considerably limits the waste of resources.

Additional experiments have been driven to measure the impact θπ and θv. As depicted in Fig. 8, high values of θπ improve the performance at high loads and when additional offset is provided. However, for low loads, a high θπ is a handicap. θv has a limited impact on the performance. The capabilities of the ADR scheme for various offered loads and various θπ values are illustrated by Fig. 9. For ρ<3, values of θπ ≤0.7 lead to the best performance, while for higher loads, θπ has to be ≥ 0.7 to reach the best carried load. Thus we select the value θπ = 0.7 hereafter.

B. EON topology

Simulations have also been performed on the EON topology under similar conditions, except that, in this case, the σ parameter takes the values 0, 1, 2 or 4 and that θπ is fixed to 0.7.

As in the SIMPLE topology case, ADR is compared to both SP and DR approaches. Performances are represented in Fig. 10. For high loads, SP does not only provide better performances than DR, but also than ADR.

Two reasons account for that. Firstly, half of the nodes of the EON topology are peripheral. They emit anyway the same traffic than other central nodes. Thus, the core part of the network quickly becomes a bottleneck as load increases. The SP scheme, by trying only once to cross the bottleneck, drops the bursts earlier, which spares resources for other bursts.

To explain the second reason for SP outperforming, the burst losses over time have been plotted in Fig. 11, for ρ=4. Three phases can be distinguished. During the first 16 ms, several losses are due to exhausted offset times. ADR, without knowing where to forward particular bursts, causes many of them to finish their lives far from their destination. Due to these frequent mislaid bursts, network resources are highly utilised. The number of burst dropped for unavailability is therefore high, too. After this first phase, ADR learns from its mistakes and does not mislay bursts anymore, for the next 65ms. On the contrary, it starts to block bursts instead of forwarding them. This blocking spares the resources, which contributes to lower the unavailable links.

After 90ms of simulation, the number of lost bursts increases again and the number of blocked burst falls, which leads to an explosion of the total losses. This is due to the fact that core node memories are limited to 2000 cells of 40µs each, giving a total memory time of 80ms.

Upon expiration of this delay, memories will start to override the information contained in their first cells, causing a progressive “amnesia”. This obliges the network to relearn the routing information.

Thus, ADR underperforms SP because it must deduce repeatedly the network structure, and needs to lose bursts in the network for this purpose. In the less complex SIMPLE topology, the network was easy to apprehend. In the EON, by contrast, many routes have to be excluded prior to find the right paths.

C. Improvement of ADR

To avoid the aforementioned handicap, the ADR scheme described in subsection III-C is modified in the following way. A forwarding option is still considered unfavourable if the values stored by the corresponding TSFC do not match the θπ and θv thresholds. However, a second requirement is added. A burst with final destination d and remaining offset r is forwarded to the output port j if and only if

ShortestPath(j,d) ≤ r+1

This condition guarantees that a burst will be forwarded only if it carries enough offset units to flow from the output terminal node to its destination. The same mechanism than DR is thus implemented. We call this modified scheme Adaptive Restricted Deflection Routing (ARDR).

In Fig. 12, the analysis of the burst drops over time permits to visualise the effect of the route restriction. Using ARDR, during the learning phase, the number of blocked burst increases steeper since fewer trials are required before starting to block burst. The number of burst dropped due to their insufficient remaining offset is oblivious null with ARDR. ARDR is also affected by the amnesia effect, but relearning is again achieved quicker due to the additional restriction. On the top lines, one can see that ARDR generally leads to a reduction of the number of losses.

Finally, the comparison depicted in Fig. 10 has been reproduced, substituting ARDR to ADR in Fig. 13. The ARDR outperforms the SP approach for σ=0 and leads to similar performances for σ=1.

1 2 3 4 5 6

offered load

ADR, σ=0

ADR, σ=1

ADR, σ=2

ADR, σ=4

DR, σ=0

DR, σ=1

DR, σ=2

DR, σ=4

Figure 10. Performance of shortest path routing (SP), deflection routing (DR) and Adaptive Deflection Routing (ADR) in terms of carried load

using the EON topology.

5016 90 150 200 250 3000

Time in ms

No availabilities

Blocked

Offset exhausted

Figure 11. Lost bursts along simulation time.

0 50 100 150 200 250 3000

Time in ms

ARDR - Total

ARDR - No availabilities

ARDR - Blocked

ARDR - Offset exhausted

ARD - Total

ARD - No availabilities

ARD - Blocked

ARD - Offset exhausted

Figure 12. Comparison of the ARDR and ARD

1 2 3 4 5 6

offered load

ADR, σ=0

ADR, σ=1

ADR, σ=2

ADR, σ=4

DR, σ=0

DR, σ=1

DR, σ=2

DR, σ=4

Figure 13. Comparison of the ARDR scheme with SP and DR routing approaches, on the EON topology

V. CONCLUSIONS

In this paper we propose an Adaptive Deflection Routing scheme, able to achieve simultaneously the operations: deflection routing, load balancing, and admission control. ADR performs all its operations by processing and analysing the feedback messages exchanged by core nodes. Nodes do not require any configuration, since they learn from past experience. Thus, the overhead required by ADR (TSFC, decisions mechanisms) is compensated by its simplicity.

ADR performance has been analysed by numerical simulation using the tool JAVOBS. Without extensively trying to deflect bursts impossible to dispatch, ADR clearly outperforms the DR scheme, especially for high loads. However, ADR has one disadvantage. It requires a minimal number of burst losses to be able to correctly apprehend the network architecture. This is problematic when low loads are injected or over networks including a large number of nodes.

To mitigate this penalty in large networks, a modified version of the ADR scheme, the Adaptive Restricted Deflection Routing has been set up. This improvement reduces the number of burst losses happening in the learning stage. Contrarily to ADR, ARDR outperforms the SP on the EON large topology. Unfortunately, the routing restriction requires the core nodes to be aware of the network topology, since it requires the knowledge the shortest paths for each source-destination pair.

The approach consisting in performing routing, deflection and admission control decisions relying only on feedbacks, without prior knowledge of link utilization, appears to be valid. This permits to achieve all these operations directly at the OBS core nodes, and spares an additional overlaid control plane. However, further analysis must be conducted, in particular regarding the convergence of both ADR and ARDR techniques. Alternative ways of storing feedback, as well as other ways to use them, could also be investigated.

REFERENCES

[1] C.-F. Hsu, T.-L. Liu, N.-F. Huang, “Performance Analysis of Deflection Routing in Optical Burst-Switched Networks”, IEEE INFOCOM, 2002.

[2] C. Qiao, M. Yoo, “Optical Burst Switching (OBS) – a new paradigm for an optical internet”, J. High Speed Network vol. 8, no. 1, pp. 69-84, 1999.

[3] S. Yao, B. Mukherjee, S. J. B. Yoo, S. Dixit, “A Unified Study of Contention-Resolution Schemes in Optical Packet-Switched Networks”, J. of Lightwave Technology, vol. 21, no. 3, 2003.

[4] Y. Chen, J. S. Turner, P.-F. Mo, “Optimal Burst Scheduling in Optical Burst Switched Networks”, J. of Lightwave Technology, vol. 25, no. 8, 2007.

[5] X. Wang, H. Morikawa, T. Aoyama, “Burst optical deflection routing protocol for wavelength routing WDM networks,” SPIE Opt. Netw. Commun. Conf., 2000.

[6] A. Zalesky, H. L. Vu, Z. Rosberg, E. W. M. Wong, M. Zuckerman, “Stabilizing Deflection Routing In Optical Burst Switched Networks”, J. on Selected Area in Communications, vol. 25, no. 6, 2007.

[7] H. Tanida, K. Ohmae, Y.-B. Choi, H. Okada, “An Effective BECN/CRN Typed Deflection Routing for QoS Guaranteed Optical Burst Switching”, IEEE GLOBECOM, 2003.

[8] M. Klinkowski, D. Careglio, J. Solé-Pareta, “Reactive and proactive routing in labelleed optical burst switching networks”, IET Commun., vol. 3, no 3, pp. 454-464, 2009.

[9] J. Zhang, H.-J. Lee, S. Wang, X. Qiu, K. Zhu,Y. Huang, D. Datta, Y.-C. Kim, B. Mukherjee “Explicit Routing for Traffic Engineering in Labeled Optical Burst-Switched WDM Networks”, ICCS, LNCS 3038, Springer, 2004.

[10] M. Klinkowski, M. Pióro, D. Careglio, M. Marciniak, and J. Solé-Pareta, “Non-linear Optimization for Multipath Source-Routing in OBS Networks”, IEEE Communications Letters, vol. 11, no. 12, 2007.

[11] P. Pedroso, J. Solé-Pareta, D. Careglio M. Klinkowski, “Integrating GMPLS in the OBS Networks Control Plane”, IEEE ICTON 2007.

[12] G. Thodime, V.Vokkarane, J. Jue, ”Dynamic Congestion-Based Load Balanced Routing in Optical Burst-Switched Networks”, IEEE GLOBECOM 2003.

[13] L. Yang, G. N. Rouskas, “Adaptive Path Selection in OBS Networks”, J. of Lightwave Technology, vol. 24, no. 8, 2006.

[14] D. Ishii, N. Yamanaka, I. Sasase, “Self-learning route selection scheme using multipath searching packets in an OBS network,” J. Opt. Netw. Vol. 4, no. 7, pp. 432-445, 2005.

[15] S. Ganguly, S. Bhatnagar, R. Izmailov, C. Qiao, “Mutli-path Adaptive Optical Burst Fowarding”, IEEE HPSR 2004.

[16] X. Gao, M. A. Bassiouni, “Fairness-Improving Adaptive Routing in Optical Burst Switching Mesh Networks”, IEEE ICC 2008.

[17] H. Pan, T. Abe, Y. Mori, Y.-B. Choi, H. Okada, “Feedback-based Load Balancing Routing for Optical Burst Switching Networks”, IEEE Asia-Pacific Conference on Communications, 2005.

[18] S. Lee, K. Sriram, H. Kim, J. Song, “Contention-Based Limited Deflection Routing Protocol in Optical Burst-Switched Networks”, J. on Selected Area in Communications, vol. 23, no. 8, 2005.

[19] J. Perelló, S. Spadaro, J. Comellas, G. Junyent, “Burst Contention Avoidance Schemes in Hybrid GMPLS-enabled OBS/OCS Optical Networks”, ONDM 2009.

[20] T. Coutelen, H. Elbiaze, B. Jaumard, “An Efficient Adaptive Offset Mechanism to Reduce Burst Losses in OBS Networks”, IEEE GLOBECOM 2005.

[21] F. Farahmand, Q. Zhang, J. P. Jue, “A Feedback-Based Contention Avoidance Mechanism for Labeled Optical Burst Switching Networks”, Photon. Netw. Commun., vol. 14, no. 3, pp. 307-316, 2007.

[22] Y. Chen, H. Wu, D. Xu, C. Qiao, “Performance Analysis of Optical Burst Switched Node with Deflection Routing”, IEEE ICC 2003.

[23] M. Klinkowski, D. Careglio, J. Solé-Pareta, and M. Marciniak, “Performance Overview of the Offset Time Emulated OBS Network Architecture”, J. of Lightwave Technology, vol. 27, no. 14, 2009.

[24] O. Pedrola, M. Klinkowski, D. Careglio, J. Solé-Pareta, S. Rumley, C. Gaumier, “JAVOBS: A Flexible Simulator for OBS Network Architectures”, J. of Networks, vol. 5, no. 2 pp. 256-264, 2010.

[25] J. Teng, G. N. Rouskas, “A detailed analysis and performance comparison of wavelength reservation schemes for optical burst switched networks”, Photon. Netw. Commun, vol. 9, 2005.

[26] B. Mikac, R. Inkret, A. Kuchar, “Advanced infrastructure for photonic networks, extended final report of Cost Action 266, ISBN 953-184-064-4.

[27] Z. Rosberg, H. L. Vu, M. Zuckerman, J. White, “Performance analyses of optical burst-switching networks”, IEEE J. on Selected Areas in Communications, vol. 21, no.7, pp. 1187-1197, 2003.

Sébastien Rumley received the M.S degrees in Communication Systems for the Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland, after studies in Lausanne, Zurich and Santiago de Chile. Since 2005 he is with the Laboratoire de Télécommunation of EPFL. His research focuses on software engineering of network simulators and planners, applied to optical networks. Christian Gaumier received his Ph.D. from Ecole Polytechnique Fédérale de Lausanne (EPFL) in 1995. His doctoral research focused on modelling the propagation of signals over singlemode fibres in both linear and nonlinear regimes. Since 2001 he is director of the Telecommunications Laboratory of EPFL. His active research area includes dispersion compensation techniques, measurement techniques for fibre optics and dimensioning and performance analysis of core photonic communication networks. He participated to several European joint projects and is author and co-author of more than 50 publications in conferences, journals and books. Dr. Gaumier is member of Communication Society of IEEE. Oscar Pedrola received the M.S. degree in Telecommunications engineering and the M.S. degree in information and communication technologies both from the Universitat Politècnica de Catalunya (UPC), Barcelona, Spain in 2008. Currently, he is a PhD student at the UPC, in the Broadband Communications Research Group (CBA). At present, he is involved in the FP7 Network of Excellence BONE. His current research interests are in the field of optical networks with emphasis on burst/packet based switching technologies. Josep Solé-Pareta obtained his M.Sc. degree in Telecom Engineering in 1984, and his Ph.D. in Computer Science in 1991, both from the Technical University of Catalonia (UPC). In 1984 he joined the Computer Architecture Department of UPC. Currently he is Full Professor with this department. He

did a Postdoc stage (summers of 1993 and 1994) at the Georgia Institute of Technology. He is cofounder of the UPC-CCABA (http://www.ccaba.upc.edu/). His publications include several book chapters and more than 100 papers in relevant research journals (> 20), and refereed international conferences. His current research interests are in Nanonetworking Communications, Traffic Monitoring and Analysis and High Speed and Optical Networking, with emphasis on traffic engineering, traffic characterization, MAC protocols and QoS provisioning. He has participated in many European projects dealing with Computer Networking topics.

Design and Development of a Semantic Information Modelling Framework for a Service

Oriented Optical Internet

Chinwe E. Abosi, Reza Nejabati, Dimitra Simeonidou High Performance Network Group

School Of Computer Science and Electronic Engineering, University of Essex, Colchester, UK, CO4 3SQ

Email: {ceabos,rnejab,dsimeo}@essex.ac.uk

Abstract— The evolution of IT and network technologies has generated far-reaching opportunities for complex service innovation in future Internet. This has resulted in an increase in the heterogeneity and complexity in service provisioning. Such complex services demand efficient coordination of distributed IT resources (storage and computing) interconnected by high capacity optical networks. This paper proposes a service plane architecture as an architectural enhancement promising to handle these complexities. It implements a unified service provisioning concept that can adapt to the heterogeneous, dynamic and complex nature of emerging service requirements as well as optical network and IT resource capabilities. The main elements of the proposed architecture consist of information discovery and service discovery that require a semantic modelling framework to address heterogeneity and automation. This paper focuses on a novel semantic modelling framework which is central to the proposed service plance architecture. It is used for information description in a service-oriented environment based on Web Services Modelling Ontology (WSMO). The framework describes the information model and entities needed to communicate requirements for autonomous, homogeneous service discovery, selection and composition.

Index Terms— service plane, description language, web services, semantic web, WSMO, future Internet

I. INTRODUCTION

The rapid advancement of future Internet applications has been motivated by the recent developments of the optical network which provides large amounts of cheap, readily available bandwidth while ensuring global reach. These future Internet applications, such as digital cinema and ultra high definition video streaming, are characterised by exchange of massive amounts of data, high levels of inter-activity, remote high definition visualisation, intensive distributed computations, and high capacity distributed storage. They address a large number of users who require dynamic access to computational, storage, visualisation and optical network resources. They are supported by various heterogeneous Infrastructure providers who expect optimised use of their resources while increasing benefits by accepting and

satisfying as many requests possible. This environment has brought about challenges in provisioning and management of resources. It has prompted the need for a framework that can efficiently support the remote, distributed and heterogeneous nature of resources as well as the dynamicity, efficiency, reliability and predictability required by applications.

To address these challenges, and ensure efficient service delivery that achieves a balance between Infrastructure providers and application needs, there is a need for a flexible, unifying and scalable platform to deliver services over the Internet [1]. Service oriented architectures have proven to be scalable, flexible means to discover, compose and deliver services. It is easy to extend and offers a good level of automation in service discovery while promising efficient resource utilisation [2], [3]. In service-oriented frameworks, resource functionalities are modelled as service elements and mechanisms are created to dynamically access and utilise these service elements efficiently in such a way as to provide the best trade-off between resource utilisation and the quality of Service (QoS).

We proposed one such service oriented framework when we introduced the service plane architecture [1]. Under this framework, the service plane architecture is responsible for exploiting the design principles of service orientation by transforming the traditional network architecture into a service-oriented architecture. It achieves this by providing a platform through which application developers and infrastructure providers can interact closely, with the purpose of providing an optimised service in which both application and Infrastructure providers’ needs are met. We define a service as a set of functionality provided by Infrastructure providers to satisfy application requirements. The unified service provisioning concept needs to tackle the heterogeneous, dynamic, distributed and complex nature of emerging application requirements and resource environments. The approach undertaken by the service plane to address these issues is by pooling information about resources from multiple Infrastructure providers into a globally accessible ‘service-oriented’ marketplace.

To manage the heterogeneous nature of resource-information in this marketplace and to facilitate autonomic and automatic discovery of resources, a common, unified vocabulary for describing resource-state information is required. Description of resource information in a formal, structured and unified manner is useful for (a) resource management, through the description of resource configurations, for example, topology and traffic flows (b) service provisioning (c) performance analysis (d) infrastructure maintenance.

In this paper, we focus on developing the formal infrastructure resource description for the purpose of service provisioning. Hence, only the components of the network and IT resources relevant to service provisioning are modelled. Ontological engineering and semantic descriptions are central to achieving this solution. Ontologies are metadata schemas that provide a structured vocabulary of information to describe data in a machine understandable format [4]. Semantics give well defined meanings to the information model and the entities needed to communicate the requirements for service provisioning using the specific background information of ontologies [5]. Thus, ontologies and semantic descriptions provide homogeneity, interoperability, consistency, knowledge formalisation, and intelligent inference. Furthermore, they facilitate precise and customised matching of service requests to available resources.

The contribution of this paper is to present a rich descriptive semantic framework and common vocabulary for the description of network and IT resources. This proposed semantic framework is based on the Web Services Modelling Ontology (WSMO) framework [6] and addresses the heterogeneity and diversity of resources, as well as the dynamicity and scale of applications which characterise the service-oriented optical future Internet. The WSMO framework allows for semantic and QoS matching using the Web Services Modelling Execution Environment (WSMX) [7]. WSMO is based on the Web Service Modelling Language WSML [8] to describe the ontologies for resources and requests, as well as their interfaces and capabilities. WSML provides the syntax and formal semantics for WSMO.

For ease of reference, we define the following concepts: • Service: a set of functionalities provided by

Infrastructure providers to satisfy application requirements.

• Service Elements: individual functionalities belonging to the complete service. Each service element satisfies a different requirement of the overall service.

• Service Discovery: the automatic location of services and resources offered by Infrastructure providers.

• Service Topology: a topology created by the composition of service elements that satisfies an application quality of service requirement.

• Virtual Service Topology: a service topology created using service elements that have been virtualised (virtual service elements).

• Web Service: an abstract representation of (virtualised) service elements which defines and describes (a) the capability of a service; (b) the interface for interacting with the service and other web-services.

The rest of this paper is structured as follows. In Section 2, the service oriented optical Internet architecture that implements the proposed semantic resource description language is presented. The relation and position of the description language within the architecture is highlighted. Section 3 discusses some related work in the area of resource description. It also introduces the WSMO model and framework. The WSMO-based Semantic Resource Description Language is introduced and discussed in section 4. Section 5 concludes the paper.

II. SERVICE PLANE ARCHITECTURE

The service-oriented optical Internet is facilitated by a service plane. The service plane is as a third party entity responsible for the provisioning and management of services. It was introduced and detailed in [1]. Fig. 1 shows the architectural scenario of the service plane.

Infrastructure providers abstract (step 1) and publish their resources into a registry as web-services, described using ontologies (step 2). The registry is composed of: (a) knowledgebase which consists of information about the resource domain (b) web-services which interface the described resources to the physical resource and other web-services (c) marketplace which consists of the instances of Infrastructure resource. Step 2 is made possible by the abstraction mechanism and semantic resource description language of the service plane architecture. Through virtualisation techniques, the infrastructure is partitioned to form virtual service-

��

� !��"��

�# �!�"��$

��%"&��#� �!�

��"��'�

(�)#�!�"��*

��'�

��+"�

��,�

��' "��!�

Figure 1: Architectural Scenario

specific infrastructures, called virtual service elements, also represented as web-services (Step 3). A user makes a query to the service plane which in turn queries the available web-services in the registry (step 4). If a match is found, it is returned (Step 5a) else a suitable service is composed (Step 5b). The result returned (Step 6).

The units involved in this scenario are depicted in Fig. 2:

A. Information Discovery

The information discovery unit is responsible for retrieving resource-state information from the optical network and IT Infrastructure providers and client request for services. This information is retrieved and managed by three main mechanisms: resource abstraction, information description, and virtualization.

The resource abstraction mechanism hides technical details of the structure of individual resources to address resource-providers confidentiality issues. The resource abstraction includes a resource-state update mechanism. The update mechanism dynamically captures resource state information and ensures that the service plane has sufficiently accurate information such that the information between the physical resources and the abstracted/logical resources are adequately consistent.

The information description mechanism addresses heterogeneity between Infrastructure providers’ and applications through the exploitation of ontologies and semantics. A novel semantic resource description language, SRDL provides a common vocabulary for the formal representation of the network and IT resource entities within the resource domain to ensure a common understanding between the entities involved in the provisioning of services. This paper focuses on describing the proposed ontology developed for this purpose. The information received from this unit is stored into a registry.

The virtualization mechanism performs service-oriented resource virtualisation by using the semantically enhanced resource descriptions to form virtual infrastructures, each optimised for a specific application.

Using virtualisation, network and IT resources are partitioned to form multiple parallel dedicated virtual resources. Each partition is dedicated for an application-type while the physical infrastructure is shared among all application-types. On the applications’ end, virtualisation portrays the illusion that the services received comprise dedicated resources optimised for their needs in terms of quality of service and functionality. On the providers’ end, virtualisation helps to improve the utilization of resources.

B. Service Discovery

The service discovery unit is responsible for discovery of existing services (IT and Network services) as well as the composition of these services (IT+Network service) to form service topologies. The service discovery unit views the virtual resources formed by the service-oriented resource virtualisation engine as virtualised service elements which can be composed into complex services. The service elements in the service discovery unit are described and accessed as web services. It comprises of three main mechanisms: 1-matchmaking engine, 2- service composition engine and 3-service adaptation.

The match-making mechanism locates services that can satisfy an application request by querying the registry. Structured, homogeneous descriptions are a critical component of the matchmaking framework. The matchmaking engine uses a matchmaking algorithm to match requests sent by the requester to the available resources offered by the Infrastructure providers. The input of the algorithm is (a) requests and (b) the resource instances stored in the registry. The output of the algorithm is the description the resources that can satisfy the request.

The service composition is invoked by the match-making engine in the case that the match-making engine is unable to find a suitable match. The service composition engine selects suitable atomic service elements and composes and connects them into complex service elements which form virtual service topologies. In the context of this paper, a web service is an abstract representation of virtualised service elements which describes the capability of a set of coherent and logically related services. If the composition engine is unable to find suitable service elements, it invokes the matchmaking engine to return its result.

The service adaptation mechanism monitors the performance of each virtual service topology as well as the physical resource-state information. Individual service topologies may be redeployed and service elements recomposed in the case of failure or performance degradation.

The service plane obtains resource information via existing management plane entities and individual monitoring and discovery services equipped in the local resource management system (LRMS) of IT resources, and the Network Management System, NMS, or the control plane of network resources. It makes reservation for the resources via User-Network Interface, UNI, (in the case of GMPLS), NMS or LRMS such as the Grid middleware (in the case of IT resources).

��

� ��

��

�� !��

��

��"��

��#$ �%��

Figure 2: Architecture

III. RELATED WORK

This section reviews relevant research. First, related work in the area of description languages is presented. Then the specification that forms the basis of the proposed description language is introduced.

A. Description Languages

A number of description languages have been proposed to describe optical networks and IT resource environments. These languages, which aim to describe the resources in complex environments, have evolved from symmetric syntactic resource description schemes to semantic description schemes. Symmetric, syntactic resource description schemes [9], [10], [11] require the properties of resources and requests to be described using flat symmetric attributes that are similar. In this type of resource description, the infrastructure provider and application developers have to be in agreement on attribute names and values. This makes this scheme inflexible, impractical, un-scalable and difficult to develop and extend [12]. Semantic schemes declaratively describe resources and requests based on expressive ontology languages. The nature of semantic description languages eliminates tight coupling between resources and requests and imposes the loose coupling characteristic of service orientation. Consequently, the addition of vocabularies and inference rules about resources and applications can be realised without the need to re-design the systems that use the descriptions. This makes it easily extensible and adaptable to changes in Infrastructure provider and application technologies [12]. We discuss the three most common underlying specifications for semantic description languages.

The Resource Description Framework (RDF) Schema [13] is an ontology modelling language and a structured basic language framework based on triples. It forms the underlying specification for the Network Description Language (NDL) [14], [15]. It expresses relationships between two entities, the subject and object, through predicates. It also allows for inference through class-subclass relationships. Inference is made through an RDF query languages, such as SPARQL [16]. However, it has limited facilities for expressing meaning and semantics which is desired for more accurate resource and service discovery.

The Web Ontology Language (OWL) [17] builds on RDF Schema by describing relationships between classes explicitly through the use of logic to make deductions about relationships. Thus it has a higher capability to share information between systems. However, its expressiveness limits its ability to automate the discovery of resources. OWL-S [18] was introduced as an extension to OWL to automate the discovery and invocation of services. OWL and OWL-S are used in several Grid Resource Description Languages [19], [20], [21]. However, OWL-S was not designed to focus on particular applications, which is important for modelling real domains. It also has limited extensibility inherited from the shortcomings of OWL [22].

WSMO [6] is the latest initiative for semantic descriptions. It is based on logical reasoning, formalisation and expressiveness similar to OWL-S with added elements to make it more extensible than OWL-S and applicable to real-domains [22]. WSMO forms the underlying specification of description language proposed in this paper.

At present, VxDL [23] is the only language that offers the opportunity for the full representation of both optical network and IT resources, such as routers, switches, latency and bandwidth as well as IT resource clusters. However, it does not explore the use of semantics that provides a high degree of flexibility and expressiveness. The NDL is a description language that uses semantics to describe network resources in hybrid networks. However, it has currently not been designed to describe IT resources.

To the authors’ best knowledge, this paper proposes the first semantic resource description language that flexibly and expressively offers a full representation of both network and IT resources.

B. The Web Services Modelling Ontology (WSMO)

WSMO is a European initiative in the area of Semantic Web Services. It describes the relevant aspects necessary for the automation of service discovery, composition and invocation. It is presented in WSML, an ontology language particularly developed for WSMO, and based on logical reasoning, formalisation and expressiveness. WSMO comprises of Ontologies, Web Services, Goals and Mediators.

• Ontologies provide a formal definition of the information model for all the other three components of WSMO.

• Web-services define and describe the capability and interfaces of the infrastructure made available by Infrastructure providers. The capability of a web service describes the functionality of the set of services it offers. The interface of a web-service describes the means through which the web-service interacts with the users of the web-service.

• Goals represent the description of the objectives an application wants to achieve. It defines the expected response to its inputs as well as the expected state after the execution of the service.

• Mediators define means to address additional heterogeneity within components of the WSMO framework. It presents a means to map resources described in other semantic languages. For this paper, it is assumed that all components within the WSMO framework are described homogeneously; hence mediators are outside the scope of this paper.

IV. SEMANTIC RESOURCE DESCRIPTION LANGUAGE

A. Aim and Scope

The definition and discovery of a desired service requires close interaction between the application and the Infrastructure provider. However, as the resources in the

future Internet environment are heterogeneous, geographically distributed with diverse users, there is a communication gap that limits the interaction between these entities. The aim of the Semantic Resource Description Language (SRDL) is to diminish this limitation and ensure efficient communication between all parties involved.

The scope of the SRDL is to specify, capture and model the entities related to service provisioning within the future Internet infrastructure. It describes the capability and interfaces to access the resources as well as application requests. The following sections introduce the design of the SRDL.

B. Structure of the SRDL

The SRDL has several parts: The common ontology, the SRDL web-service based on the WSMO web-service, and the SRDL goal based on the WSMO goal.

The SRDL common ontology describes the important aspects of IT and optical network resources related to service provisioning. It consists of several component ontologies which are designed separately to model the resource domain. The entities and vocabulary used to build the ontology relate to the resources they represent.

The SRDL web-service imports relevant aspects of the common ontology to form instance ontologies that describe the constraints over the functionality provided by each infrastructure provider. The web service, through the instance data and the resource interface, provides data to the marketplace.

The SRDL goal, like the web-service imports relevant aspects of the ontology to instantiate its request. The SRDL goal queries the registry for appropriate SRDL web-services.

The common ontology together with the instance ontologies of the SRDL web-services and goal provide the information that builds the knowledge base of the registry.

C. Modelling the Resource Domain

The resource domain is modelled to form common ontology of the SRDL. The ontologies that form the common ontology capture the important aspects of the resource domain including the QoS characteristics of resources and the terminology that define the relationships of QoS characteristics. Resources in service-oriented environments are defined as entities offered by the Infrastructure providers which are required for implementing and satisfying application requirements. In this paper, we deal with the representation of optical network, storage and computational resources. Since the application is not concerned with the technological details of the resources, but rather the capability of the resources, the SRDL abstracts out this information. The SRDL ontology is sub-divided into five sub-ontologies:

• The Resource Ontology describes the devices, interfaces and the connections that exist between them. The resource ontology starts by describing

distinctive network and IT resource entities, such as nodes, links, interfaces and paths. Additionally, it differentiates, categorises and describes different types of entities and sub-categorises them, for example, networkNode and ITNodes are types of Node. Each resource entity within the resource ontology is represented as a class (Fig. 3).

The attributes of each resource is represented as properties. The properties represent the relationship and constraints between the entities. These properties give semantic meaning to the ontology. For example, the link entity is described these attributes: connectsToNodes, hasCapacity, hasFreeCapacity, hasType (for example optical fibre), hasDirection (bi-directional or uni-directional), hasLength, hasDelay(the propagation delay), hasCost. The properties are also defined by a cardinality. The cardinality determines the number of values each property can have. Most attributes within the resource domain ontology have single cardinalities while a few, for example, the connectsToNode has a cardinality of two, since a maximum of two routers can connect to a single link.

A simplified version of the resource ontology is shown in Fig. 3. Below we give a short description of some of the classes within the resource domain ontology:

Node: A device attached to a network. NetworkNode: A node that is capable of sending,

receiving, or forwarding information over a communications channel

ITNode: A node that is used for processing or storing data.

Compute: An IT node that comprises at least one CPU capable of processing data.

Storage: An IT node that comprises at least one hard disk capable of storing data

Server: An IT node attached to a network whose primary purpose is to store application and data files to the shared. In our scenario, this is a video server.

ResType: A listing of all types of resources that exist, i.e. storage, compute, server and network

VideoFile: A file that contain information as video Interface: A device used by a node to connect to

another node in a network Link: A communication channel that transfers

data between two nodes. Path: An abstract connection between at least

two end-points in a network.

• The Instance Ontologies describes the capacities of each instance of the resources that exist within the resource domain.

• The Resource QoS Ontology describes the QoS parameters that relate to the entities described within the resource domain ontology such as DataSize and

CPUSpeed. It comprises relevant QoS parameters that profile each type of resource and application QoS.

• The QoS Base Ontology describes general QoS parameters and the relation between them. It defines the semantics of QoS parameters and forms the building block for the resource QoS ontology. For example, it describes a dataSizeUnit which has as attributes, hasphysicalQuantity and hasUnit. It also describes the relation between them through the use of axioms. Axioms improve the semantics of the ontology by refining the classes and relationships between classes and attributes. Axioms also add declarative knowledge and express constraints in the form of logical expressions. For example, axioms define the relationship between a GigaByte Unit and than a MegaByte Unit through a conversion factor. An axiom can also define comparison criteria through logical expressions. These comparison criteria allow inference to be made by the matchmaking engine to consider only offered resources with quantities that are either greater than, less than or equal to the requested quantity.

• The Adapter Ontology describes the instances and

value mappings between values in the WSMO ontology and their corresponding input and output

properties of a Web Service Description Language (WSDL) document used for communication and execution through the Simple Object Access Protocol (SOAP).

• The Service Description Ontology describes the structure of responses and requests for each of the possible requests. For example, we modelled a storageRequest with attributes, sourceNode, duration, requiredDiskSpace, requiredBandwidth. A storageResponse class has as its attributes destintationNode and confirmation. Similar classes and corresponding attributes are modelled for computational and connectivity requests and responses. The attributes describe the expected input parameters for a request and the provided output parameters returned in response to the request.

Fig. 4 shows an example of a graph representing an instance of a typical infrastructure/resource domain that comprises of three network routers that connects a storage node, a computational node and three video servers in a network. Fig. 5 shows a snippet of the SDRL representation of the resource domain instance.

ResourceOntology

Resource::Router

+hasID : int

+connectsToNode : Node

+hasFreeCapacity : Capacity

+hasCapacity : Capacity

+hasDirection : Direction

+hasType : string

+hasLength : Distance

+hasCost : double

+hasDelay : Delay

Resource::Link

+hasFreeDiskspace : DataSize

+hasTotalDiskSpace : DataSize

+hasFreeMemorySize : MemorySize

+hasTotalMemorySize : MemorySize

+hasCPUModel : CPUModel

+hasCPUSpeed : CPUSpeed

+hasOS : OS

Resource::ComputeNode

+hasFreeDiskspace : DataSize

+hasTotalDiskSpace : DataSize

Resource::StorageNode

+hasID : string

+hasCapacity : Capacity

Resource::Interface

+hasName : string

+hasFileSize : DataSize

Resource::File

+hasID : int

+hasLocation : Location

+hasType : NodeType

+hasInterface : Interface

Resource::Node

Resource::ITNode Resource::NetworkNode

+hasFile : File

Resource::Server

Resource::NodeType

+hasName : string

Resource::Location

Resource

Resource::Switch

Resource::Optical Resource::Electronic

QoSOntology

QoS::Quality

QoS::CPUSpeed

QoS::Direction

+hasPhysicalQuantity : double

+hasUnit

QoSBase::Metric

QoSBase::CPUSpeedUnit+hasConversionFactor : double

QoSBase::QoSParamUnit

Figure 3: Classes and Attributes of a Resource Domain

D. Modelling the Offered Infrastructure

The resources that each Infrastructure provider offers are viewed as atomic services offered by this provider, which will be combined and composed to address application requests as long as enough resources exists. These resources are published and exposed as WSMO web-services. These web services are defined by their capabilities and interfaces through which they interact with other services or a requestor of a service. The capability of a web service describes the functionality of the set of concrete services it represents (Fig. 6).

Web-services are described using: • Preconditions or the expected input. For example,

a source, required disk space and bandwidth must be specified to access a storage web service.

• Assumptions: conditions assumed for a service to be executed: the duration specified must be greater than zero.

• Postconditions: the desired output a suitable storage node has been identified based on the constraints given. A confirmation has also been given for this request.

webService WSProvider ... capability WSProviderCapability ... preconditiondefinedByexists

?storageReq(?storageReq[ sd#hasSource hasValue ?srcid, sd#reqDiskSpace hasValue ?reqdisk, sd#reqFreeCapacity hasValue ?bwcap ] memberOf sd#StorageRequest). postconditiondefinedByexists

?storageResp(?storageResp[ sd#hasDestination hasValue ?dest, sd#confirmation hasValue ?confirm ] memberOf sd#StorageRequest). effectdefinedBy isReduced(StorageRequest).

interface WSProviderInterface importsOntology {...} choreography WSProviderChoreography stateSignature WSProviderStateSignature in concept sd#StorageRequest withGrounding _"http://www.examples.org/storage?wsdl" out concept sd#StorageResponse transitionRules _# forall{?request} with (?request memberOf sd#StorageRequest) do add(_#1 memberOf sd#StorageRequest) endForall

Figure 6: Example Web Service representing the Offered Infrastructure by a provider as a WSMO Web-Service

ontology WSProviderDQoS nonFunctionalProperties wsmostudio#version hasValue "0.8.0" endNonFunctionalProperties importsOntology {...} //Optical Networkinstance Router1 memberOf ro#Router ro#hasID hasValue 1 ro#hasType hasValue ro#Router ro#hasLocation hasValue ro#Location_1 instance Router2 memberOf ro#Router {...} instance Router3 memberOf ro#Router {...} instance Link1 memberOf ro#Link ro#hasID hasValue 1 ro#connectsToNode hasValue {Router1, Router2 } ro#hasDelay hasValue ro#Delay_1 ro#hasLength hasValue ro#R1_to_R2 ro#hasDirection hasValue ro#Bidirectional ro#hasFreeCapacity hasValue _ro#capacity_link1 ro#hasType hasValue ro#optical_fibre instance Link2 memberOf ro#Link {...} instance Link3 memberOf ro#Link {...} //Video Servers instance Server1 memberOf ro#Server ro#hasID hasValue 1 ro#hasType hasValue ro#Server ro#hasInterface hasValue ro#interface_10Gig ro#hasFile hasValue rio#File1 ro#hasLocation hasValue ro#Location_3 instance Server2 memberOf ro#Server {...} instance Server3 memberOf ro#Server {...} //Storageinstance Storage1 memberOf ro#Storage ro#hasID hasValue 1 ro#hasType hasValue ro#Storage ro#hasInterface hasValue ro#interface_10Gig ro#hasDiskSpace hasValue sio#storageDiskSpace3 ro#hasLocation hasValue ro#Location_2 //Computeinstance compute1 memberOf ro#Compute ro#hasDiskSpace hasValue cio#diskSpace1 ro#hasMemorySize hasValue cio#memorySize1 ro#hasID hasValue 1 ro#hasInterface hasValue ro#interface_10Gig ro#hasType hasValue ro#Compute ro#hasCPUSpeed hasValue cio#cpuSpeed1

ro#hasLocation hasValue ro#Location_1

Figure 5: Snippet of the Instance ontology for above network

Figure 4: A resource configuration representing a graph of a typical infrastructure

• Effect: of service execution: available bandwidth and disk space on selected route and storage device respectively have reduced.

The interface of the web service describes under what conditions a web service or a client can access and interact with the web service capability. The transition rule describes how to get the response to a request. The transition rule in Fig. 8 specifies that all requests made have to receive a response. WSMO web services are described unambiguously using the terminology defined in the resource domain ontologies.

E. Modelling the Requests

Clients are presented with a form-based interface to describe their requirement for IT and network resources. This interface transparently integrates semantic and non-semantic items of a request. The request is parsed by a request handling unit which constructs a formal service request based on WSMO Goals. WSMO defines a goal template at design time and a goal instance at run-time. A goal instance is a goal template instantiated with a client’s request. This process removes the need for a client to have pre-existing knowledge of the information model or domain ontologies used to describe services. Goals are defined using postconditions and effects only. The goal is passed to the service discovery unit.

Fig. 7 shows the template of a storage request. The goal specifies the desired capability for each of the QoS characteristics of network and IT resources it requires through the hasEnoughStorageRes() and storageConnected() logical expressions. Fig. 8 illustrates these logical expressions..

The logical expression checks for storage nodes with enough QoS parameters to satisfy the request. It then tries to find a path to this node. A result is only retrieved in the case that there are enough storage resources as well as network bandwidth capacity connecting the requesting node (source node) to the discovered node (destination

node). In the case where more than one solution is discovered for a given request, applications can specify which QoS characteristic is preferred in selecting a service when multiple solutions are discovered for a given request. This function limits the optimality of the utilisation of resources; however, it gives requesters the opportunity to fine-tune the results of their request. However, in this work, preference is given to the first service discovered is used.

F. The SRDL Design Architecture

We provide a scenario that depicts the overall design architecture of the SRDL in Fig. 9. • Information from Infrastructure providers are parsed

into SRDL web-services and published into a repository.

• Application requests are parsed using the WSMO goal to form SRDL goals. The requests specify the characteristics of the resources required, for example, a request for a storage node that has a storage capacity

goal GoalReqStorageA capability GoalReqStorageACapability ... postcondition

definedBy sd#hasEnoughStorageRes(storageRequest) and sd#storageConnected(connectionRequest)interface GoalRequestStorageAInterface importsOntology

{GoalRequestA...} choreography GoalReqStorageAChor stateSignature GoalReqStorageASS in concept sd#StorageResponse out concept sd#StorageRequest transitionRules GoalATransitionRules forall {?request} with (?request memberOf sd#StorageRequest) do add(_#1 memberOf sd#StorageResponse) endForall

Figure 7: WSMO Goal requesting Storage Resources

ontology GoalRequestA axiom hasEnoughStorageResDef definedBy?storageRequest[ sd#reqDiskSpace hasValue ?reqDisksize] memberOf sd#StorageRequest

and ?reqDisksize [qosbase#value hasValue ?reqdisk] memberOf qos#DiskSpace and

(?MyStorage[ ro#hasID hasValue ?theStorID, ro#hasDiskSpace hasValue ?myDisksize, ro#hasLocation hasValue ?cpuloc] memberOf ro#Storage and ?myDisksize [qosbase#value hasValue ?mydisk] memberOf qos#DiskSpace)

and ?reqdisk =< ?mydisk implies sd#hasEnoughStorageRes(sd#storageRequest)

axiom storageConnectedDef definedBy

?connectionRequest[ sd#reqFreeCapacity hasValue ?reqcapacity, sd#reqSource hasValue ?reqsrcid] memberOf sd#StorageRequest

and ?reqcapacity [qosbase#value hasValue ?reqcap] memberOf qos#Capacity

and (?MyPath[ ro#hasFreeCapacity hasValue ?myCap, ro#hasDelay hasValue ?myDelay, ro#hasID hasValue ?linkID, ro#connectsWith hasValue{?srcnode,?destnode}] memberOf ro#Link and ?myCap [qosbase#value hasValue ?mycap] memberOf qos#Capacity and ?srcnode[ro#hasID hasValue ?srcid] memberOf ro#Router and ?destnode[ro#hasID hasValue ?destid, ro#hasLocation hasValue ?destloc] memberOf ro#Router )

and wsml#numericEqual(?reqsrcid, ?srcid) and ?reqcap =< ?mycap and wsml#stringEqual(?destloc, ?cpuloc)

implies sd#storageConnected(sd#connectionRequest)

Figure 8: Logical Expression to Query the Repository to satisfy a Storage Requirement

of 20GB. The parsed request or SRDL goal uses its interface to query the repository for resources based on the capability of both the SRDL goal and the SRDL web-services.

• The registry is searched for potential web-services that can satisfy the goal that can satisfy the application requirements.

• The result is retrieved and invoked.

CONCLUSION

In this paper, we presented a service plane architectural extension based on service orientation as a possible solution for emerging Internet services. The paper focused on a novel semantic information description framework based on WSMO. The semantic resource modelling framework proposed here is the first that offers a full representation of both network and IT resources and their attributes, such as routers, switches, latency and bandwidth as well as computing clusters, storages and application servers. It is able to provide uniform access to network and IT resources as well as client requirements, which prevents conflicts and misinterpretation brought about by heterogeneity. It also enables efficient, autonomous service discovery and composition required by service oriented network environment.

ACKNOWLEDGMENT

The work described in this paper was carried out with the support of the European Commission through the GEYSER (“Generalised Architecture for Dynamic Infrastructure Services”) project in the Seventh Framework Programme.

REFERENCES

[1] C.E. Abosi, R. Nejabati, and D. Simeonidou, "A novel service composition mechanism for future optical Internet", IEEE J. Opt Comm and Netw, vol 1, pp. A106-120, July 2009.

[2] V. A. S. M. de Souza, and E. Cardozo, “SOANet - A service oriented architecture for building compositional network services”, J. Soft., vol. 1. pp. 1-11, August. 2006

[3] T. Erl, Service-Oriented Architecture (SOA). Concepts, Technology, and Design. Prentice Hall PTR, 2005.

[4] M. Obitko, "Ontologies Description and Applications", Technical Report, Czech Technical University in Prague, Prague, 2001.

[5] T. Berners-Lee, J. Hendler, and O. Lassila, “The Semantic Web”, Scientific American, Vol. 284, pp. 34-43, May 2001.

[6] W3C Web Service Modelling Ontology (WSMO), W3C Member Submission, 3 June 2005 [http://www.w3.org/Submission/WSMO/. [accessed online 27 January 2010]

[7] W3C Web Service Execution Environment (WSMX) W3C Member Submission, 3 June 2005 http://www.w3.org/Submission/WSMX/. [accessed online 27 January 2010]

[8] W3C Web Service Modelling Language (WSML) W3C Member Submission 3 June 2005 http://www.w3.org/Submission/WSML/ [accessed online 27January 2010]

[9] R. Raman, M. Livny, and M, Solomon. “Matchmaking distributed resource management for high throughput computing”, IEEE Int Symp. High Per., Dist. Comp, pp. 140-146 July 1998.

[10] C. Liu, and I. Foster, “A constraint language approach to matchmaking”, Proc 14th Int W on Res Iss on Data Eng: Web Serv for E-Comm and E-Gov App,. pp. 7 -14, March 2004.

[11] A. Chien,. “The virtual Grid description language: vgDL”, Technical Report, University of Carlifonia, San Diego, August 2009.

[12] H. Tangmunarunkit, S. Decker, and C. Kesselman, “Ontology-based resource matching in the Grid - The Grid meets the semantic web” LNCS: The Semantic Web ISWC, Springer, October 2003, pp. 706-721.

[13] W3C RDF Primer. W3C Recommendation 10 February 2004 http://www.w3.org/TR/rdf-primer [accessed online 27 January 2010].

[14] J.J. van der Ham, F. Dijkstra, P. Grosso, R. van der Pol, A. Toonk, C.T.A.M. de Laat, "A distributed topology information system for optical networks based on the semantic web", J. Opt. Switch. and Net., vol 5 pp. 85-95 June 2008.

[15] J. van der Ham, F. Dijkstra, F. Travostino, H.M. Andree and C. de Laat, "Using RDF to describe networks", Fut. Gen. Comp. Sys., vol. 22 pp. 862-867, October 2006.

[16] W3C SPARQL Query Language for RDF W3C Recommendation 15 January 2008 http://www.w3.org/TR/rdf-sparql-query/ [accessed online 27 January 2010]

[17] W3C OWL Web Ontology Language Overview, W3C Recommendation 10 February 2004. http://www.w3.org/TR/owl-features [accessed online 27 January 2010]

[18] W3C OWL-S: Semantic Markup for Web Services, W3C Member Submission 22 November 2004, http://www.w3.org/Submission/OWL-S/ [accessed online 27 January 2010]

(��!��

��

�(-.

/��

��

(�)#�!�

��0��

��

�!��-��

��

�!��-��

�(-.

��

(�!�#��

��0��

(�!�#��1��0��

�# �!�

-�!��!

(�)#�!�

��

2#��

3!�!

Figure 9: SRDL Design Architecture

[19] S. Ludwiga, and S.M.S Reyhani, “Semantic approach to service discovery in a Grid environment” J. Web Sem., vol. 4, pp. 1-13, January 2006.

[20] D. Trastour, C. Bartolini, and J.A. Gonzalez-Castillo, “Semantic web approach to service description for matchmaking of services”. Proc. Int. Semantic Web Working Symp., July 2001.

[21] S. Kailash, P. Prasanna, V. Prabha, V. Neela Narayanan, "Semantic resource description for Grid" in Proc First Asia Int. Conf. on Model. & Simul., pp. 112-115, March 2007.

[22] R. Lara, D. Roman, A. Polleres, and D. Fensel. “A conceptual comparison of WSMO and OWL-S”. In Proc. of the 2nd European Conference on Web Services, 2004.

[23] G. Koslovski, P. V. Primet, and A. Schwertner Charão. VXDL: Virtual resources and interconnection networks description language. In Proc GridNets 2008, October. 2008.

Chinwe E. Abosi is a PhD candidate at the School of Computer Science and Electronic Engineering at the University of Essex, Colchester, United Kingdom. She became a student member of IEEE in 2006. She received her BEng degree in electrical and electronics engineering from the University of

Botswana in 2005, and her MS degree in information networking from Carnegie Mellon University, Pittsburgh, Pennsylvania, USA, in association with the Athens Information Technology Centre, Athens, Greece in 2006.

Reza Nejabati joined University of Essex in 2002 and he is currently a member of Photonic Network Group at the University of Essex.

During the last 8 years he has worked on ultra high-speed optical networks, service oriented and application-aware networks, network service virtualization, control and management of optical networks, high-performance network architecture and technologies for e-science. Reza Nejabati holds a PhD in optical networks and an MSc with distinction in telecommunication and information systems from University of Essex, Colchester, United Kingdom.

Dimitra Simeonidou is currently a professor at the University of Essex. She has over 10 years experience in the field of optical transmission and optical networks. In 1987 and 1989 she received a BSc and MSc from the Physics Department of the Aristotle

University of Thessalonica, Greece and in 1994 a PhD degree from the University of Essex.

From 1992 to 1994 she was employed as Senior Research Officer at University of Essex in association with the MWTN RACE project. In 1994 she joined Alcatel Submarine Networks as a Principle Engineer and contributed to the introduction of WDM technologies in submerged photonic networks. She participated in standardisation committees and was an advising member of the Alcatel Submarine networks patent committee.

Professor Simeonidou is the author over 250 papers and the holds 18 patents relating to photonic technologies and networks. Main research interests include optical wavelength and packet switched networks, network control and management and Grid networking.

Reducing Complexity and Consumption in Future Networks

G. M. Tosi Beleffi, G. Incerti, L. Porcari, S. Di Bartolo, M. Guglielmucci, Ministry of Economic Development, Communication Department, ISCTI, Viale America 201, Rome, Italy email:

giorgio.tosibeleffi@sviluppoeconomico.gov.it and gabriele.incerti@sviluppoeconomico.gov.it, lucaporcari3@virgilio.it, silvia.dibartolo@sviluppoeconomico.gov.it, michele.guglielmucci@sviluppoeconomico.gov.it

A. L. J. Teixeira, L. Costa,

Instituto de Telecomunicacoes de Aveiro, Portugal email teixeira@ua.pt, costa@ua.pt

N. Wada,

National Institute of Communication and Information Technologies Tokio, Japan email: wada@nict.jp

J. Prat, J. Lazaro,

Universidad Politecnica de Cataluna, Barcelona, Spain jprat@tsc.upc.edu, lazaro@tsc.upc.edu

P. Chanclou

France Telecom R&D, 2 avenue Pierre Marzin, 22300 Lannion, (France) philippe.chanclou@orange-ftgroup.com

Abstract: The authors report the main results of the EU FP7 SARDANA project, a future network architecture able to provide large bandwidth to the end user in a flexible and intelligent way. The remote amplification for network reach extension, the remote monitoring of the network infrastructure, and the possibility of remotely powering ONUs are fundamental capabilities to maintain the passiveness of the network outside plant, so important for reducing power consumption and maintenance costs, without compromising network reach and coverage.

Index Terms: passive optical networks, carbon footprint reduction, remote functionalities, WDM/TDM technology

I. INTRODUCTION In the past few years optical networks experienced, as a long wave effect, a substantial increase in the transported bandwidth in all the sectors, from the high speed transport segment to the low speed access up to the end user. At the event horizon, the introduction of mixed fully transparent network architectures [1], ring+tree, to cover all the metro-access scenario, open the way to the implementation of extended broadband metro networks from the core up to the end user. In this scenario, it has been experienced a boost in the transmitted bandwidth/data rate, more and more close to the end user premises, and a massive introduction of passive plants in the access segment to cope with the green challenges and to lower the capital expenditures (CAPEX). The potential deployment of 40G and 100G links, furthermore, implies significant challenges even in well-controlled high performance networks (core like). These high-speed signals are naturally more influenced to propagation related impairments such as

chromatic dispersion (CD) and polarization mode dispersion (PMD) and nonlinearities. Furthermore, in Next Generation Access Network (NGAN) scenarios dominated by long reach passive plants, and in the general context of the global crisis & global warming, it becomes important to monitor the physical integrity of the fiber infrastructures, to push the deployment of green technologies and applications with limited cost increase. GPON and GePON standards, for example, fix working wavelengths at 1480 nm for down stream (DS) and 1300 nm for upstream (US). Different wavelengths are under study but still have not been implemented. Remote transparent devices able to translate standard transmission bandwidths (C- and L- bands) to standard PON working wavelengths, and vice versa, could become crucial to avoid an increase of end-user expenses and to fully open the network to non incumbent operators. Having this capability at the physical layer will surely result in new business models and opportunities, giving the fiber plant operators and deployers the opportunity to directly exploit them without requiring data from the second layer network users. Centralizing key functionalities like for example amplification and wavelength conversion, or adopting simultaneous remote measurements of impairments, such as power variations, optical signal-to-noise ratio (OSNR), CD, PMD and physical infrastructure integrity, on existing networks, becomes a crucial point to speed up the decisions around moving from 10G to 100Gbit/s deployments. Last but not least, the chance to supply the end users with voltage free optical network terminations (ONTs) is a challenge that could open the way for a new

paradigm in the telecom segment. In this paper it is given a general view of the current network scenario, with particular emphasis on next generation Metro-Access infrastructures. The main results of the EU FP7 SARDANA [1] and BONE Projects [2] are also presented, in terms of remote functionalities like amplification, monitoring, signaling and powering.

II. REDUCING THE COMPLEXITY, THE SARDANA INFRASTRUCTURE

The SARDANA network transparently merges TDM single-fiber passive tree sections with a main WDM double-fiber ring by means of passive Remote Nodes (RN), as shown in Fig II.1.

Fig.II.1. The SARDANA Network

The 100km WDM ring transports 32 wavelengths for >1000 users, for a TDM tree splitting ratio of 1: 32 and only 1 wavelength per TDM tree. Network protection and traffic balancing properties are provided by the ring configuration and the resilient design of the RNs, guarantying always a connection between each RN and the Central Office (CO) even in the case of fiber cut. The WDM ring is implemented by a double-fiber to avoid main Rayleigh backscattering (RB) impairments. Bidirectional propagation takes place at the single-fiber TDM trees. The wavelength transparency of the ring and ONUs, and the wavelength add/drop feature of the RNs enables sharing of the same network infrastructure, until the RNs or even up to the ONT/ONU, by several operators and allows users to select the operator by easily exchangeable filters at the ONU. This way, SARDANA offers a possible solution for implementing multi-operability in the physical layer, by simply allocating a set of wavelength DS/US channels to each operator. The WDM/TDM overlay in SARDANA network eases the migration process from legacy PON solutions, such as standard G/EPON , and next–generation 10G versions. Future standard 10G/E-PON OLT&ONT can be adapted to the SARDANA PON by the corresponding optical interfaces [6]. At the same time, SARDANA aims to reduce the complexity of the network infrastructure by implementing a fully passive plant from the CO to the end user premises, with several RNs interconnecting the WDM ring with the TDM trees.

Fig.II.2. SARDANA network deployment scenarios

The RN amplifies the signals flowing in upstream (US) and downstream (DS) by means of remotely pumped erbium doped fiber (EDF) samples. In order to evaluate the real beneficial effects to brand new as well as to pre installed infrastructures, several scenarios have been taken into account. See Fig. II.2.

Fig.II.3. Simplifying the network can reduce both investment and

maintenance costs, as well as power consumption

In Fig.II.3 is reported one of the main results demonstrating the easiness of this approach to simplify the network architecture reducing costs, both CAPEX and OPEX. By simply adopting an intelligent OLT location with both a capillary fiber optic diffusion and the implementation of a regional/sub regional SARDANA infrastructure would be, in principle, possible to reduce the number of CO from 820 to 11, thus having a tremendous impact on the carbon footprint, as well as on the operator maintenance costs reduction. This case study has been conducted in the Bretagne area.

III. REMOTE AMPLIFICATION FOR METRO ACCESS CONVERGENCE

The RN becomes a key element of the ring + trees network. It, in fact, encompasses key challenges, like passiveness in the sense of not requiring electrical power supply, efficient 1480 nm pump use, and burst mode amplification avoiding gain transients. The inset in Figure.III.1 shows a RN with 2 drop wavelengths (one for each TDM tree) and bidirectional remote amplification by means of EDF sections. Wavelength extraction is done by means of two athermal thin-film OADMs at alternated 100 GHz or 50 GHz ITU-T grid channels. The natural gain transients due to the amplification of US dynamic burst-mode are cancelled in the RN thanks to the simultaneous amplification of both Continuous Wave (CW) DS and US channels in the same EDF section, also avoiding RB of the remote optically pumped amplifier. The implemented RN presents 1 dB insertion loss in by-pass, 6 dB in drop/add, and > 30 dB rejection. The losses are largely compensated by about 14 dB gain of the EDF. In [7] this is compared to other types of Extender Boxes for PONs, in terms of reachable trunk & access power budget. We specify up to 16 RNs, thus 32 wavelength channels, with a splitting ratio between 1 and 32 each.

Fig.III.1. Network test configuration with 105 km ring, RN scheme (inset) and DS (above) /US (below) transmission BER

measurements.

First tests of the SARDANA network have been performed, in different configurations. Fig. III.1 shows the scheme and results in a 105 km ring between Rome and Pomezia cities, at 10G DS and 2.5G US, with 2 RNs and 3 channels. The pump power at the CO was below 1.2 watts at 1480nm. Sensitivities are -33 and -36 dBm respectively. Protection against fiber cut was validated, with less than 1 dB penalty at rerouting, in DS and US directions. In order to test the importance of the Raman gain presence along the ring, a rural scenario has been used. The fiber ring is 8.9km with 8 RNs and 16 WDM channels. In each RN, the add/drop amplification of two channels is used and optimized for minimum required pump power from CO. The add/drop amplification gain is set to obtain the required signal power at ONU input, which is -15dBm required for the

remodulation of the US by DS signal achieved by reflective SOA. The tree fiber is 20km and the splitting ratio is 1:16, giving a tree loss of about 17dB. The Raman amplification and noise are calculated by Matlab tool for each ring span as a function of the signal and pump power, the wavelength, the noise power at the fiber input, and the number of channels. The obtained simulation results are verified experimentally for 80km of total ring fiber distance and with reconfigurable RN [6]. All pump required is transmitted together with the DS signals in co-propagating direction. Higher RN pass loss is used, corresponding to the two extra WDM multiplexers in each RN. For these calculations, the required pump power for each RN is used to give the optimum add/drop amplification without the influence of the Raman effect. The considered pump power at CO output is 1W at 1480nm. The considered noise power is -60dB at CO output for DS signals.

Fig.III.2. OSNR as a function of the signal wavelength for the

signal power of a) 0 and b) 10dBm. The dashed line with circle mark stands for simulation result achieved without Raman amplification,

solid line with rhombus mark stands for simulation and square marks stand for experimental analysis achieved with Raman amplification.

In Figs.III.2 a) and b), an overall OSNR as a function of signal wavelength for 0 and 10dBm of signal power are shown, respectively. The results of only four and five RNs which are closer to the CO (8 and 10 dropped channels) are shown, where Raman gain is higher. About 8dB of the overall OSNR improvement with Raman plus add/drop amplification compared to only add/drop amplification scheme can be seen in Fig.III.2, independently of the signal power.

IV. REMOTE AND WAVELENGTH TUNABLE MONITORING

As reported in Fig.IV.1, the monitoring system is composed by a standard Optical Time Domain Reflectometer (OTDR) from Wavetek (a), a 50%

coupler, an Optical/Electrical (O/E) converter (b) with an amplifier stage (c) able to drive, with the correct voltage levels, an External Cavity Laser (ECL) (d) . The pulsed optical signal generated by the Wavetek OTDR is wavelength converted by an o-e-o conversion stage. The backscattered signal is received by the Wavetek OTDR photodiode after being directed with an optical circulator.

Fig.IV.1. Set-Up

The position of a reflection is evaluated on the basis of the time necessary to the light to return to the OTDR, assuming that forwarding and back-warding paths are the same. There are two different paths in this system due to the O/E conversion. The pulse delay depends on various system parameters, but in the configuration used, it corresponds to around 110ns which is approximately equivalent to a position evaluation error of about11m, and thus to a negligible distance for usual distance range in the order of km. With an emitted signal equal to -8,7dBm (peak power), the best performance has been found to be for a nominal pulse length of 1 µs, that gives a real pulsewidth, i.e. after the ECL of 1.12 µs. This little discrepancy is mainly due to the presence of a high pass inverter present in the ECL electronic modulation stage and not reported, for simplicity, in Fig.IV.1. The inverter is needed to avoid an inverted logic at the ECL output with respect to the OTDR output. By using a better inverter, with a lower pass cut frequency, the system would be able to follow the OTDR when a longer pulsewidth is set. The experimental set up is reported in fig. IV.2.

Fig.IV.2 Set-Up. Gr1 Fiber Bragg Grating at 1554nm. Gr2 Fiber

Bragg Grating at 1557nm.

The wavelength tunable OTDR is placed in the CO and connected to a SMR fiber with 11.8km via a 2x2 3M PLC coupler. The fiber coil represents the network feeder fiber. The goal of the system is to discriminate different Fiber Bragg Gratings (FBGs), nominally Gr1

and Gr2, placed along the GePON network by changing the ECL wavelength. The SMR fibre coil is interconnected to a 3M PLC 2x32 network splitter via a FBG (Gr1) with central wavelength at 1554nm; after the 2x32 splitter, 3 ONTs AN5006 (Optical Network Termination Units) from FiberHome are connected. The 2x2 coupler insertion loss is equal to 3.2 dB, and the 2x32 splitter insertion loss is equal to 15.57 dB at 1550nm. In front of one of them has been placed an SMR fibre coil 1km long and a FBG (Gr2) with central wavelength at 1557nm. The Wavetek OTDR has been set to generate pulses at 1550 nm, 1µs long and with a distance range of 80km. The OLT has been switched on in order to test the monitoring tunable system in a real active network.

Fig.IV.3. (A) OTDR Wavelength at 1557nm (B) OTDR

wavelength at 1554nm.

In Fig.IV.3 (A), is reported the OTDR trace when the ECL generates pulses at 1557nm. As it can be seen, the grating is perfectly located. Thus it demonstrates that, from the CO, the operator can simply check the status of the network even after the high attenuation induced by a 2x32 splitter. In principle, by inserting 32 gratings in front of each ONT is possible to monitor the status of the entire access network discriminating each path by a simple wavelength tuning. In Fig.III.3 (B) the ECL wavelength is set at 1554 nm, corresponding to the gr1 central wavelength. In this case, there is a great intensity peak just after the feeder section. At this wavelength the Gr2 does not produce a reflective event. The OLT and the ONT traffic have not been affected by the presence of the monitoring signal.

V. REMOTE POWERING OF THE ONT Last but not least, preliminary results on a new kind of cabling fiber are reported. A Fijikura low friction cable with G652A bend insensitive optical fibre has been connected to an ONT via standard voltage supply. The experimental set up is reported in Fig.IV.2. The ONT is a standard ePON ONT device from Fiber Home. It needs 12V and 25W to operate. The special cable is made by a G.697A optical fiber and two steel wires as strength member (see fig.V.1). The sheath is done by a low friction and abrasion resistant flame retardant polyolefin jacket. This reduces the friction and the installation time. Fujikura developed this new cable in order to solve the problems due to the usage of guide ropes, and the presence of several indoor cables in vertical cabling. Preliminary results on a two meter

long Fujikura cable demonstrate the possibility of remotely serve the ONT, thus opening the window to the implementation of zero supply voltage network termination units.

Fig.V.1 Preliminary remote powering experiments.

This is an inherent energy saving for the end user. The ONT in fact, switches on itself sending packets to the OLT. Ongoing experiments are being performed to check the QoS of the entire system, from the OLT to the ONT in order to understand how this special kind of remote voltage supply affects in some ways the network performance.

VI. CONCLUSIONS It has been shown the main results of the SARDANA EU FP7 project. This architecture, which is fully passive and long reach is able to provide all the main

characteristics of the Future Networks, i.e. large bandwidth per end user, flexibility, scalability, carbon footprint reduction, low CAPEX and OPEX.

ACKNOWLEDGMENTS

This work is supported by the EU FP7-ICT project "SARDANA" (G.A.: 217122). Authors want to thanks also Fujikura and NewFont Italia S.r.L.

REFERENCES

[1] www.ict-sardana.eu [2] www.ict-bone.eu [3] C. C. Lee, T. C. Kao, and S. Chi, “Simultaneous

Optical Monitoring and Fiber Supervising for WDM Networks Using an OTDR Combined With Concatenated Fiber Gratings”, IEEE Photonics Technology Letters, Vol. 13, No. 9, pp. 1026 – 1028, 2001.

[4] Y. C. Chung, “Optical Monitoring Techniques for WDM Networks”, Electronic-Enhanced Optics, Optical Sensing in Semiconductor Manufacturing, Electro-Optics in Space, Broadband Optical Networks, 2000. Digest of the LEOS Summer Topical Meetings, pp. IV43-IV44, 2000.

[5] S. Hann, J. Yoo and C. Park, “Monitoring technique for a hybrid PS/WDM-PON by using a tunable OTDR and FBGs”, IOP Meas. Sci. Technol., Vol. 17, pp. 1070–1074, 2006.

[6] A. Baptista et al, ECOC’2008, paper Tu1F6. [7] F. Saliou, P. Chanclou, F. Laurent, N. Genay, J. A.

Lazaro, F. Bonada and J. Prat. “Reach Extension Strategies for Passive Optical Networks", J. OptCom. Networks, vol.1, no.4, 2009.

Decrease of the Link PMD by Fiber Exchange and Investigation of the PMD Distribution along

Buried Optical Fibers with a POTDR A. Ehrhardt

Deutsche Telekom AG, Technology Engineering Center, Goslarer Ufer 35, 10589 Berlin, GERMANY Email: A.Ehrhardt@telekom.de

M. Paul*, L. Schürer, C.Gerlach, W. Krönert**, D. Fritzsche****, D. Breuer***, V. Fürst*****, N.

Cyr******, H. Chen******, G.W. Schinn*****Deutsche Telekom Netzproduktion GmbH, Zentrum TE, Goslarer Ufer 35, 10589 Berlin, Germany

*Deutsche Telekom AG, GHS, Goslarer Ufer 35, 10589 Berlin **Deutsche Telekom AG, TI NL Süd, Schürerstraße 9a, 97080 Würzburg, Germany

***Deutsche Telekom T-Labs, Goslarer Ufer 35, 10589 Berlin, Germany ****European Center for Information and Communication Technologies – EICT GmbH, 10587 Berlin, Germany

*****T-Systems Enterprise Services GmbH, Deutsche-Telekom-Allee 7, 64295 Darmstadt, Germany ******EXFO Electro-Optical Engineering Inc., 400, av. Godin., Québec QC G1M 2K2, CANADA

Email: {Manuel.Paul, Lars.Schuerer, C.Gerlach, Wolfgang.Kroenert, D.Breuer}@telekom.de; Daniel.Fritzsche@eict.de; Volker.Fuerst@t-systems.com, ncyr@total.net, {hongxin.chen, greg.schinn}@exfo.com

Abstract— The accumulated PMD along several fibers in an optical network was measured in order to identify those short fiber portions exhibiting high PMD. By replacing high-PMD-portion(s) in each fiber, the overall-PMD could be reduced significantly. Additionally the PMD-accumulation of different fibers in the same buried cable was investigated and compared to each other. Index Terms— PMD, OTDR, 40G/100G transmission, field trial, optical fibers, backbone fiber network

I. INTRODUCTION

The fiber network of most network operators generally comprises a heterogeneous mixture of fibers and cables, produced by different vendors over many different years. Thus, there can be a correspondingly important variation in the physical properties of deployed fibers. Along with fiber attenuation and chromatic dispersion, polarization mode dispersion (PMD) is an important parameter, as it can limit data rates and transparency length.

During the last years Deutsche Telekom has reviewed its network [1] and characterized the PMD value of a large number of fibers. Fiber links exhibiting excessively high PMD values for transmission of 40 Gbit/s and beyond were identified. A single high-PMD section can be sufficient to render an entire 400 km link unusable for high-speed transmission. In order to avoid such data rate and transmission length limitations, PMD compensators can be used for each channel [2], an approach that can become quite costly and unwieldy. Alternatively, or in addition, advanced modulation schemes (e.g., 40G RZ-DQPSK) can be applied to partially mitigate the sensitivity to PMD [3]. The third possibility to reduce the influence of link PMD is to identify and replace the “bad”

fiber sections [4]. Whereas a replacement of all fiber sections along an entire link is extremely costly and time consuming. Fortunately, it is generally the case that the square of the accumulated PMD along a fiber link does not grow linearly as a function of distance, but rather that this growth tends to be localized in one or just a few distinct portions or sections accounting for the large majority of the total link PMD. As a result, an identification and selective replacement of those few distinct portions, enabled by the availability of a suitable measurement technique is much more cost effective and could be economically competitive to the deployment of PMD-tolerant transmission systems.

In this paper, a prototype tunable random-scrambling polarization optical time-domain reflectometer (RS-POTDR) [5] was used in a field trial to identify and exchange fiber sections along buried cables having high PMD in the network of the fixed network division of the Deutsche Telekom group (T-Home). In order to compare the obtained results and demonstrate the successful improvement, the cumulative PMD along the fibers was characterized before and after replacement.

II. MEASUREMENT PRINCIPLE AND PROCEDURE

A classical OTDR measures the attenuation along an optical fiber by emitting short optical pulses and performing time-gated detection of the Rayleigh-backscattered light (or reflection from localized Fresnel reflections) of each pulse as a function of time delay (optical distance) along the fiber. In this way, localized information on the fiber attenuation, splice and connector quality can be obtained. It has long been “standard procedure” to use an OTDR to locate and characterize fiber sections showing excessive loss and/or connectors

exhibiting high reflection, and, if necessary, repair or replace them in order to ensure that the transmitted signals correspond to predetermined physical limits. However, as mentioned above, ensuring low attenuation is not sufficient for links carrying high data rates over long-haul distances, since the signal quality is also very much limited by the link PMD. Although overall (“end-to-end”) PMD can be characterized for a link with one of a number of different established measurement techniques, up to now, there has not been a proven reliable means to locate and quantify “bad PMD” sections in a field-installed link.

Using the prototype RS-POTDR instrument, a cumulative PMD curve can be obtained as a function of fiber length and high-PMD fiber portions thereby identified [5]. The wavelength of the laser will be tuned and for each wavelength the input state of polarization (SOP) is randomly scrambled. A screenshot of a typical measured cumulative PMD curve for a test fiber is shown in Fig. 1. A traditional OTDR trace corresponding to the backscattered power is overlaid on the same distance scale. The fiber portions contributing most significantly to the cumulative PMD can be identified. The detailed optical setup of the RS-POTDR prototype used for this work is described in [4].

III. RESULTS OF MEASUREMENT OF FIBERS WITH IINCREASED PMD

A first trial to test the measurement principle under field conditions was performed in the network of Deutsche Telekom [4]. Fibers were analyzed and portions having high PMD values could be identified. The main goal of this field test was to identify the bad fiber sections of a fiber link in a cable, to repair the cable and to verify the results by repeating the measurements after the repair.

PMD distribution of fiber 3: Σ PMD 5.65 ps

Fig. 1: Original PMD distribution of fiber 2 and fiber 3

We used a cable in which one fiber exhibited a high PMD value. This fiber was the “fiber under test” (FUT) which had to be repaired. Another fiber having a low PMD value served as a “donor” fiber to be grafted into and replace that section of the FUT which significantly contributed to the PMD. Fiber#3 with an overall PMD value of 5.65 ps was the FUT requiring repair and fiber#2 with an overall PMD value of 0.56 ps served as the donor fiber. In Fig. 1 the local PMD-distribution for both fibers before replacement is shown. As a check, we repeated the measurements from the opposite end and confirmed an accurate concordance of both the overall PMD and the corresponding local distribution. Additionally the overall PMD values of both fibers were characterized using the interferometric method (TINTY), i.e. without local resolution. These independently obtained overall-PMD values were in good agreement with those from the RS-POTDR, within a maximum deviation of 10 %.

IV. MEASUREMENT RESULTS AFTER FIBER EXCHANGE

It is anticipated that the most likely approach to improve the PMD behavior of the fiber infrastructure will involve actually replacing the entire “bad” cable portion contributing preponderantly to the cumulative PMD. Unfortunately we had no opportunity to do this in our field trial, and it is for this reason that we used the “donor fiber”. As seen in Fig. 2 the section between 2.0 km and 4.0 km of fiber#3 must be replaced. In this fiber portion fiber#3 has a PMD value of approximately 5.4 ps. Fortunately, fiber#2 exhibits a low contribution to the cumulative PMD value in this section. Thus fiber#2 is an ideal donor fiber for fiber#3 in order to decrease its

End point A End point B

PMD after exchangeFiber 02: 6.01 psFiber 03: 0.41 ps

PMD before exchange:Fiber 02: 0.56 psFiber 03: 5.65 ps

Total cable length: 19.03 km2 km 4 km

Fibers Fiberscable sleeve

End point A End point B

PMD after exchangeFiber 02: 6.01 psFiber 03: 0.41 ps

PMD before exchange:Fiber 02: 0.56 psFiber 03: 5.65 ps

Total cable length: 19.03 km2 km 4 km

Fibers Fiberscable sleeve

Fig. 2: Exchange of fiber portions for replacement and emulation of a cable repair

overall PMD value and to verify the measurement principle. The topology of the replacement is shown in Fig. 2. The measured PMD curves were correlated with the map of the fiber junction boxes, thus permitting the identification of the appropriate junction boxes to be opened for the cable exchange. We opened one fiber junction box located approximately at km 4 and replaced the first 4 km of fiber#3 by fiber#2 and vice versa. In this way, we anticipated a significant decline of PMD of fiber #3 and additionally that the newly-modified fiber#2 would now exhibit a “bad” fiber section, corresponding to that formerly of fiber#3.

Fig. 3: PMD distribution of fiber 3 and fiber 2 after exchange

In Fig. 3 the cumulative PMD along the fibers was characterized after interchange of the first 4 km in both fibers. The PMD value of fiber#3 decreased to 0.41 ps (versus 5.65 ps before) and fiber#2 had a PMD value of 6.01 ps (versus 0.56 ps before). Thus fiber#3 which previously was not suitable for high-speed data transmission was now improved by substitution of the short portion of fiber. Additionally the 4-km high-PMD section moved to fiber#2, thereby clearly demonstrating the underlying field suitability of the measurement principle.

V. INVESTIGATION OF THE PMD-DISRIBUTION OF DIFFERENT FIBERS IN THE SAME CABLE

In the previously described field experiment the basic suitability of the measurement principle for field deployment was tested. The next step was to investigate the PMD distribution of all free fibers with increased accumulated PMD in a common buried cable to make these fibers suitable for high speed transmission systems. The economically best way for that [6] is to replace simultaneously the sections with increased PMD values for all fibers to be improved. Therefore these fibers had to be measured and the obtained results with the PMD distribution of all fibers had to be analyzed and compared to each other. Two different cables with a number of free fibers with lengths of 31.5 km (cable A-B) and 31.08 km (cable C-D) respectively were measured and the PMD distribution of these fibers was recorded. Additionally the splice maps with the location of the sleeves were added to the measured PMD curves. In Fig. 4 the PMD distribution of the investigated fibers in the cable A-B is depicted. In a bar chart the local PMD values of the different sections

which are divided by the sleeves are shown as a function of the fiber length in the cable.

Cable A - B

0 5 10 15 20 25 30 35Fiber length [km]

Fiber 11Fiber 12Fiber 19Fiber 23Fiber 30Fiber 39Fiber 40Fiber 20

Cable A - B

0 5 10 15 20 25 30 35Fiber length [km]

Fiber 11Fiber 12Fiber 19Fiber 23Fiber 30Fiber 39Fiber 40Fiber 20

Fig. 4: PMD distribution of all fibers in the cable A-B

As a result of comparison of the different fibers it can be seen that the PMD along the fiber length is not uniformly distributed and that the sections with the main PMD contribution for the different fibers are in different sections. If we consider only a reduced number of measured fibers in this cable we can conclude that the PMD distribution of these smaller number of fibers can be relatively homogenous (fiber 12) and that sections mainly contributing to the total PMD value can be identified (fibers 11 and 30) as shown in Fig. 5. Furthermore the comparison of the fiber 11 and fiber 30 shows that the mainly contributing sections for these two fibers are unfortunately not located in the same section of the cable. That means that for an improvement of the PMD value of at least two fibers the replacement for fiber 11 must be done in one section of the cable and the replacement for fiber 30 must be done in another piece of the same cable. Despite of the replacement of only a short cable piece for each fiber to be improved probably a larger number of cable sections in total must be touched for the improvement of both fibers.

Cable A - B

0 5 10 15 20 25 30 35

Fiber length [km]

Fiber 11Fiber 12Fiber 30

Cable A - B

0 5 10 15 20 25 30 35

Fiber length [km]

Fiber 11Fiber 12Fiber 30

Fig. 5: PMD distribution selected fibers in the cable A-B

A similar behavior concerning the PMD distribution of different fibers in cable C-D is shown in Fig. 6. The PMD of some fibers exhibiting elevated PMD values is concentrated in one or a few short cable sections. In order to find an economically suitable approach to decrease the total PMD value it is necessary to characterize all

available fibers in the cable and to compare the individual Cable C - D

0 5 10 15 20 25 30 35Fiber length [km]

Fiber06Fiber08Fiber12Fiber13

Fiber14

Fiber16Fiber25Fiber26Fiber05Fiber15

Cable C - D

0 5 10 15 20 25 30 35Fiber length [km]

Fiber14

Fig. 6: PMD distribution of all fibers in the cable C-D

PMD distribution. All the free fibers in the cable should be carefully analyzed − not just those fibers having comparatively low PMD values. In this way, for each problematic cable section, one can usually identify one or more fiber pairs within that section that are suitable for high speed transmission systems, thereby avoiding physical replacement of that cable section.. However, even if this is not practical, replacement of only those short cable pieces responsible for the majority of the PMD can be economically reasonable instead of the replacement of the whole cable.

VI. CONCLUSION

In a field trial, the PMD distribution in deployed fiber sections was investigated using a novel POTDR prototype instrument, this instrument enabling the single-sided measurement of the cumulative PMD of optical fibers. The presented results show the individual PMD of different fiber sections. Hence, sections with very high PMD can be identified with sufficient spatial resolution, and can be subsequently replaced. The measurement technique can not only be applied to old fibers in order to replace bad pieces selectively, but can be also applied for newly installed cables directly after installation to check the quality and the adherence of the fibers to predetermined physical limits. This information is important for network operators who want to improve their networks in order to install systems at 40 Gbit/s and beyond. The measurement technique was applied to investigate the PMD distribution of free fibers in buried cables. The comparison of the obtained PMD curves has shown that the main contribution to PMD for any given fibers in a cable is often found in a different cable segment than that of another fiber in the cable. Despite of the replacement of only one short cable piece for each fiber to be improved a larger number of cable sections must be touched for the improvement of more than one fiber in the cable. But nevertheless the partial replacement of short cable pieces for different fibers on different places can be economically reasonable instead of the replacement of the whole cable.

REFERENCES

[1] A. Gladisch et al.: Evolution of Terrestrial Optical System and Core Network Architecture, Proceedings of the IEEE, Special Issue “Technologies for Next-Generation Optical Networks”, Vol. 94, No.5, May 2006.

[2] Chongjin Xie et al.: Automatic Optical PMD Compensator for 40-Gb/s DBPSK and DQPSK with Fast Changing SOP and PSP, ECOC 2008, Brussels, Belgium, September 2008, paper W3eE5.

[3] C. Fuerst et al.: 43 Gb/s RZ-DQPSK DWDM Field Trial over 1047 km with Mixed 43 Gb/s and 10.7 Gb/s Channels at 50 and 100 GHz Channel Spacing., ECOC 2006, Cannes, France, September 2006, PD-paper Th4.1.4.

[4] A. Ehrhardt et al.: Characterisation of the PMD distribution along optical fibres by a POTDR. Invited paper 10. ICTON 2008, Athens, Greece, June 22-26, 2008, paper We.A1.3, Proc. Vol.1. pp. 173-177.

[5] N. Cyr, H. Chen, and G.W. Schinn: “Random-Scrambling Tunable POTDR for Distributed Measurement of Cumulative PMD”, Journal of Lightwave Technology, Vol. 27, No. 18, September 15, 2009, pp. 4164 – 4174.

[6] A. Ehrhardt et al.: Characterisation of the PMD Distribution along Optical Fibres and Improvement of the Backbone Fibre Infrastructure by a POTDR., Journal of Networks Feb. 2010, Special Issue: Transparent Optical Networkimg, www.academypublisher.com

Armin Ehrhardt was born in Freiberg, Germany, in 1964. He received the Dipl.-Phys. degree from the Kharkov State University, Ukraine, in 1987 and the Ph.D. degree in physics from the Humboldt University Berlin in 1991 for research on WDM-systems with heterodyne receiver schemes. From 1991 to 1996 he worked on photonic components, generation of short optical pulses with semiconductor lasers and

high speed optical transmission systems at the Heinrich-Hertz-Institute Berlin. In 1996 he joint the Technology Center of Deutsche Telekom where he managed laboratory experiments and field trials with WDM transmission. Currently he is with the Technical Engineering Center of Deutsche Telekom Netzproduktion GmbH where he is responsible as a senior expert for improvement of the fiber infrastructure, measurement equipment, concepts of next generation optical networks of Deutsche Telekom and investigations on high capacity WDM transmission. He authored or co-authored more than 90 refereed and invited journal and conference papers and several patents.

Manuel Paul was born in Forst, Germany, in 1980. He obtained his Dipl.-Inf. degree in computer science from the University of Applied Sciences, Leipzig, in 2005. In 2005 he joined T-Systems as research professional and technology consultant where he has been actively involved in a broad range of projects concerning packet transport technologies and transmission systems, next

generation network design and corporate R&D in the fields of Carrier Grade Ethernet, DWDM/OTN and fiber optic network infrastructure.

Currently he is with the fixed network standardization department of Deutsche Telekom AG, GHS in Berlin, Germany, as a senior consultant. His work focuses on carrier grade packet transport elements for DT’s next generation network and corporate standardization activities.

Lars Schuerer was born in Annaberg-Buchholz, Germany, in 1977. He received the Dipl.- Ing. (FH) degree from the Deutsche Telekom Hochschule für Telekommunikation Leipzig (FH), University of Applied Sciences, Germany, in 2005 for analyses and characterisation of an optical transmission system based on Polymer Optical Fiber. From 2006 to 2008 he worked for T-Systems Enterprise Services

GmbH where he was involved in different field trails e.g. to analyse of new optical transmission systems or measurement equipment. Currently he is with the Technical Engineering Center of Deutsche Telekom Netzproduktion GmbH in Berlin where he is involved in concepts for development of next generation optical networks of Deutsche Telekom and in investigations on high capacity WDM transmission.

Christoph Gerlach was born in Leipzig in 1975. He received his Dipl.-Ing. (FH) degree from the Deutsche Telekom Fachhochschule für Telekommunikation Leipzig, University of Applied Sciences, Germany, in 2005 for analyzing DWDM and CWDM transmission system characteristics. In 2002 he joined the Broadband Network Architecture and Photonic Systems Design Group, Technology Center of Deutsche Telekom, now T-

Systems Enterprise Services GmbH, in Berlin. His main focus was on transmission system evaluation, interoperability testing, and calculations on economic aspects. Since 2009 he is with the Hybrid Network Technologies Department of the Technical Engineering Center of Deutsche Telekom Netzproduktion GmbH. Among other things he is now involved in developing concepts for deployments of NGN, mainly for the core and aggregation domain, and in energy efficiency in telco and home networks.

Wolfgang Krönert was born in Würzburg 1958. He received his technician degree (Tech. staatl. geprüft) in electrical engineering in 1982. After this he did further studies at Fachhochschule Dieburg University of Applied Science. From 1990 up to now he attended the development of the optical fiber in the field and the roll out of different services on it as a field engineer. Since 2000 he´s got also a concentration of high transmission

links of subscriber lines in the access network of the Deutsche Telekom. Beside this he works as a part-time lecturer on the Deutsche Telekom Training Center of Stuttgart.

Daniel Fritzsche was born in Germany in 1979. He received the Dipl.-Ing. degree in electrical engineering from Dresden University of Technology, Dresden, Germany, in 2004. From 2004-2008, he was with the Communications Laboratory, Dresden University of Technology working toward the Ph.D. degree in close cooperation with T-Systems Enterprise Services GmbH, Berlin,

Germany. At T-Systems, he worked on adaptive equalization techniques in high-speed optical transmission systems. In 2009 he joined the European Center for Information and Communication Technologies (EICT) in Berlin, Germany as project manager, where he is involved in research projects focused on carrier Ethernet and next generation optical access. He authored or co-authored more than 25 refereed journal and conference papers.

Dirk Breuer was born in Düren, Germany, in 1967. He received the Dipl.-Ing. and Dr.-Ing. degree in electrical engineering from Technical University of Berlin, both with distinction, in 1993 and 1999, respectively. During his Ph.D. his main research interest was in high capacity optical transmission systems with bit-rates of 10 and 40 Gbit/s, focusing on the impact of different intensity modulation formats, fiber nonlinearities and fiber types for

metro and long-haul applications. From 1998-2000 he was with Virtual Photonics Inc., responsible for the development of a physical layer WDM design tool. In 2000 he joined the Broadband Network Architecture and Photonic Systems Design Group, Technology Center of Deutsche Telekom, now T-Systems Enterprise Services GmbH, in Berlin, Germany. He was mainly concerned with developing optimization strategies for the optical transport network of Deutsche Telekom. including theoretical and experimental evaluation of new optical WDM systems and related technologies. Dr. Breuer is a member of the Institute of Electrical and Electronics Engineers (IEEE) and the German Verein Deutscher Elektrotechniker (VDE). He authored or co-authored more than 60 refereed journal and conference papers.

Volker Fürst was born in Berlin, in 1959. He received the Dipl.-Ing.(FH) degree in telecommunication technology from University of Applied Sciences, Leipzig and the Dipl.-Ing. degree in electrical engineering from Technische Universität Dresden, Faculty of Transportation and Traffic Sciences, in 1983 and 1991, respectively. From 1991 to 1995 he worked in the field of thin film hybrid

circuits and photonic components at Research Institute of Deutsche Telekom in Berlin. From 1996 to 1999 his was involved in research on ultra high speed electronics. In 2000 his scope of work changed to photonic networks. He was involved in evolution and test of the first 40Gbit/s WDM systems in the Deutsche Telekom fiber network. Currently, he is a Technology Consultant with T-Systems, Systems Integration in Darmstadt, Germany, with a focus on in the Next Generation Network program of Deutsche Telekom.

He is a part-time lecture at University of Applied Sciences, Wiesbaden, Germany where he gives lectures on system availability and network recovery.

Normand Cyr obtained his Ph.D in Physics Engineering at COPL/Laval University (Quebec, Canada) in 1990 on the subject of atomic clocks, precision metrology and interaction of polarized light with matter. He pursued his work at COPL as a Research Associate until 1997 when, after developing an interest in optical fiber measurements

through external contracts, he joined the Research Group at EXFO Electro-Optical Engineering. He is currently semi-retired but still holds the status of Principal Research Scientist at EXFO. He is the author of numerous papers and holds multiple granted or pending patents.

Hongxin Chen obtained his B.Sc. in laser physics at Department of Physics of University of Science and Technology of China in 1988, and M.Sc. in optics at Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Sciences, in 1993, where he continued as a Research Associate until 1995. He obtained a Ph.D. in physics at Department of Physics and Astronomy of The Open

University, England, in 1999 in quantum optics / laser cooling and trapping of atoms. From 1998 to 2000, he held postdoctoral fellowships in the Physics Departments of the University of Strathclyde, Scotland, and York University, Toronto, respectively, in laser cooling / atomic physics. In 2000, he joined the Re-search Group at EXFO Electro-Optical Engineering where he is now a Principal Research Scientist. He is the author of numerous papers and holds several granted or pending patents.

Gregory W. Schinn obtained his Ph.D. in Physics at JILA/University of Colorado in 1988 in atomic/laser spectroscopy. Following a postdoctoral fellow in atomic physics at the University of Viriginia, he worked as a scientist at MPB Technologies Inc., in Montreal, Canada from 1990 – 96, where he initiated the company’s activities in EDFA and fiber laser development. In 1996

he joined EXFO Electro-Optical Engineering, holding variously the positions of Scientific Director, CTO, and Director of Research/Intellectual Property. He is the author of numerous papers and holds several granted or pending patents.

Deployment and Validation of GMPLS-Controlled Multi-layer Integrated Routing over

the ASON/GMPLS CARISMA Test-bed

Fernando Agraz, Luis Velasco, Jordi Perelló, Marc Ruiz, Salvatore Spadaro, Gabriel Junyent and Jaume Comellas

Advanced Broadband Communications Center (CCABA), Universitat Politècnica de Catalunya (UPC), Barcelona, Spain

Email: comellas@tsc.upc.edu

Abstract— The efficient accommodation of sub-wavelength client flows on optical channels is a current challenge for the optimization of resources in GMPLS-controlled optical networks. While the capacity of the optical channels usually exceeds 10 Gbps, connection requests show finer granularity. This paper concentrates on the design and implementation of the ASON/GMPLS CARISMA test-bed a GMPLS-controlled grooming-capable transport network. Through the paper, the operation of a GMPLS-controlled multi-layer network architecture is introduced, reviewing those implementation issues that come into light. Then, an experimental evaluation is conducted on two alternative network scenarios with different number of nodes and nodal degree. From the results, GMPLS-controlled grooming makes a good trade-off between network blocking probability and E/O port usage when compared to all-optical and opaque solutions, leading to enhanced network performance while reducing network capital expenditures (CAPEX). Index Terms— GMPLS, multi-layer networks, sub-wavelength client flows, network test-bed.

I. INTRODUCTION Wavelength-routed optical networks have received

increasing attention as a promising approach to deploy end-to-end transparent networks in a cost-effective way. The definition of the ITU-T Automatically Switched Optical Network (ASON) architecture [1], allows wavelength-routed optical networks to include dynamic connection capability. This capability is accomplished by means of a control plane entity, responsible for the establishment, maintenance and release of connections over the optical transport plane.

In parallel, the Internet Engineering Task Force (IETF) has standardized Generalized Multi-Protocol Label Switching (GMPLS, [2]) as a set of protocols to implement a common control plane, able to manage several switching technologies in an integrated way.

GMPLS is the most widely accepted solution to implement the control plane functionalities in the ASON architecture. These ASON networks with a GMPLS-enabled control plane are typically referred as ASON/GMPLS networks.

The role of IP as a convergent technology has triggered the development of a wide range of new multimedia services, like HDTV, video conference, telemedicine applications or Internet telephony, having each one different Quality of Service (QoS) requirements. This huge, heterogeneous and predominantly bursty generated traffic poses new challenges to network operators to provide a cost-effective data transmission. Because the bandwidth granularity of wavelength-routed optical networks is very coarse, typically a whole wavelength supporting 10 or even 40 Gbps, these networks lack the flexibility to support sub-wavelength traffic demands, which leads to a poor bandwidth usage.

In this context, the term traffic grooming identifies the process of packing several low-speed traffic streams into higher-speed streams, trying to maximize optical channels bandwidth usage in Dense Wavelength Division Multiplexing (DWDM) meshed transport networks [3], [4]. From the GMPLS point of view, the grooming problem is translated into merging several higher-order Label Switched Paths (LSPs) into a lower-order LSP (e.g., grooming packet LSPs carrying IP traffic into a λ-LSP). Such an LSP aggregation in GMPLS is accomplished by advertising newly created lower-order LSPs as Forwarding Adjacency LSPs (FA-LSPs, [5][6]), for instance, by means of the OSPF-TE protocol [7]. In this way, conventional data-links (coming from wavelength channels) along with the previously advertised FA-LSPs can indistinctly enter the path computation process. Supposing that a valid route would be found, resource reservation would then be performed by RSVP-TE [8].

The goal of this paper is to introduce the design and implementation of the grooming-capable ASON/GMPLS CARISMA test-bed. To this end, section 2 introduces the

FA-LSP concept defining a routing metric and a cost function. Section 3 details the scenario where this integrated multi-layer routing has been evaluated also reviewing the implementation of the FA-LSP functionalities. Then, section 4 presents the obtained experimental results. Finally, section 5 concludes the paper.

II. MULTILAYER ROUTING THROUGH FA-LSPS From an architectural point of view, the control plane

in multi-layer networks can follow three differentiated models namely overlay, augmented and peer [6]. In the traditional overlay model different control plane instances run on each layer. Thus, they are independently controlled. Alternatively, in the augmented approach, different control plane instances run on each layer but some information is exchanged amongst them, aiming at improved bandwidth allocation in the network. Finally, in the peer approach all layers are controlled by a unified control plane, and decisions are taken considering whole network information.

The enhanced TE protocols introduced in GMPLS pave the way to peer multi-layer network architectures, controlled by means of a GMPLS-enabled common control plane. The enabling entity to this goal is the FA. In GMPLS, those already established lower layer LSPs (e.g., λ-LSPs) are advertised as FA-LSPs, which can be used to transport new sub-wavelength client LSPs. In this way, lower layer resources can be effectively utilized.

Fig. 1 shows an example of a two-layered network peer architecture, particularly, an optical server layer and a client aggregation layer on top (e.g., SONET/SDH, MPLS, GbE etc.). At the bottom, optical nodes provide DWDM ports as well as client access ports, used to inject an aggregated client flow to the network. In this context, an incoming signal would be adapted, switched to a DWDM port, multiplexed into a DWDM bundle and finally transmitted to the downstream optical node. On top, the client aggregation layer includes generic nodes providing electrical switching, flow aggregation and other features. Client nodes are connected to optical nodes through the client access ports.

In current GMPLS standardization [2] there is an intrinsic association between the signaling of new client LSPs and the creation of the required λ-LSP to support them. As will be later detailed, a route from source to

O/EPorts

OpticalNode

AggregationNode

AccessPorts

DWDMPorts Wavelength

Channels

FA-LSP

λ-LSP

DWDMOptical Link

Fig. 1 Example of a two-layered optical network where a λ-LSP has been set up. A FA-LSP with some residual bandwidth is advertised.

destination is computed upon client LSP request, which may be constituted of both unallocated data links and already existent FA-LSPs. In the case that no FA-LSP is comprised along the route, a new λ-LSP is typically set-up from source to destination to support the incoming request. Otherwise, λ-LSPs are set-up to provide connectivity on those route segments where no FA-LSP is yet established. This operation, however, may lead to resource waste in the network. Notice that long FA-LSPs connecting far-off nodes are limited to be only reused by incoming LSP requests between remote endpoints. Hence, it appears more appropriate to separate the signaling functionality from the λ-LSP creation, so that λ-LSP placement can be decided based on network characteristics.

In the present paper, we set the routing metric of the already established FA-LSPs to be max{1, hops(FA-LSP)-1}, as described in [5]. Besides, the routing metric assigned to the unallocated data links spanning one single physical hop is set to 1. Aiming at better resource utilization, however, we dissociate λ-LSP establishment from network signaling functionality in the following way. Once the route from source to destination is calculated, the heuristic cost function CFA(H) = H [(1-pH) + h/H] is applied to the route segments where connectivity is not yet existent, standing H for the number of hops of the yet to be created λ-LSP and pH for the probability that any incoming demand in the network has a certain number of hops H.

The cost function provides us with the most appropriate λ-LSP configuration to optically connect the yet uncovered route segment. As will be later depicted by example, the term (1- pH) encourage those λ-LSP lengths close to the average network distance, thus being more likely to be reused. The term 1/H identifies the use of O/E port pairs per hop, so that the larger the λ-LSP, the lower the use of O/E ports to connect its endpoints. The tunable h parameter fosters/penalizes the use of O/E ports in the network. Finally, the total cost is multiplied by H as longer λ-LSPs need a higher number of unallocated data links. In this context, let us imagine that a new λ-LSP, which will afterwards act as FA-LSP, has to be established between a node-pair distancing 4 hops. Supposing that a 2 hops client LSP length is the most likely in the network, the combination CFA(2)+CFA(2), that is, two λ-LSPs spanning each one two hops, could have lower cost than CFA(4), meaning one single end-to-end λ-LSP. Very short FA-LSP establishment (e.g., 1 hop) is also penalized in CFA(H), due to the large amount of expensive O/E ports required, as well as the huge amount of bypass traffic to be electrically processed.

III. IMPLEMENTATION OF THE GMPLS-CONTROLLED MULTI-LAYER INTEGRATED ROUTING

The experimental evaluation of the GMPLS integrated routing functionality has been carried out over the ASON/GMPLS CARISMA test-bed [11], a configurable

Fig. 2 The OCC architecture in the CARISMA test-bed: Link Resource Manager (LRM), Routing Controller (RC) and Connection Controller (CC).

multi-topology Signaling Communications Network (SCN) running over Optical Cross-Connect (OXC) emulators. In this configurable SCN, Optical Connection Controllers (OCCs) are interconnected by 100 Mbps full-duplex Ethernet links, describing the same topology than the underlying transport plane. In the test-bed, OCCs are deployed by means of Pentium IV Linux-based routers at 2~GHz, so that each OCC implements the full GMPLS protocol set: RSVP-TE for signaling, OSPF-TE for routing and information advertisement LMP for resource discovery and management [12].

Fig. 2 depicts the OCC architecture in the CARISMA test-bed, highlighting the standards followed on each composing module to achieve the desired operation. Each OCC contains three modules that interact among them: the Link Resource Manager (LRM), the Routing Controller (RC) and the Connection Controller (CC).

The LRM module is responsible for the management of the resources available at the optical node. Specifically, the state of the transport plane resources is stored in the Management Information Base (MIB). The OXC Manager module synchronizes the state of these resources with the optical node through the Connection Controller Interface (CCI). The same interface is used by the node to notify alarms using Simple Network Management Protocol (SNMP) traps [13]. Moreover, the LRM contains the LMP module that implements the GMPLS LMP protocol, and the LRM server module, which implements communication interfaces to the RC and CC modules in the same OCC.

The RC, basically, is the responsible for computing routes. It implements several routing algorithms to

compute transport plane routes. In this regard, it is worth noting that the CARISMA network test-bed uses differentiate addressing spaces at the control plane and at the transport plane. In fact, the quagga OSPF module [14] implements the OSPFv2 protocol which floods links state information related to the control plane IP network. In contrast, the RC floods the state of the local output data-links to the rest of control plane OCCs using OSPF-TE Opaque Link State Advertisements (OLSAs, [7], [15]). The information in the OLSAs is related to the transport plane and stored in the TE database (TEDB). OLSA flooding is performed every time a data-link is allocated or released. The RC module implements communication interfaces to the CC and LRM. The CC requires route computation between two end nodes, whereas the LRM notifies the RC about the reservation, release or failure of the local resources, which imply OLSA flooding.

Finally, the CC is responsible for the LSP set-up and tear-down. The CC module includes the RSVP module which implements the RSVP-TE protocol [8], [10]. The CC contains the Path State Block (PSB) database which stores every LSP supported on local resources. The Network Management System (NMS) communicates with the CC through the NMI-A interface to request set-up or tear-down connections. Upon receiving a connection set-up command towards a remote node, the CC asks the RC for a transport plane valid route to that node. Then, every CC on the route to the destination must ask the LRM about the availability of the local resources and eventually request them to be allocated.

The implementation of the FA-LSP functionality at the control plane of the CARISMA test-bed has been made

Connection Controller

CC Module

RSVP Module

NMI‐A

P‐TE

MsgFSM

CC Server

CC_LRM

Link Resource Manager

LRM Module

OXC Manager

LMP Module

MsgFSM

ConfigurationAlarms Notification

LRM Server

Routing Controller

RC Module

quagga OSPF

F‐TE

MsgFSM

RC Server

OSPF API

Algorithms

RFC 3630, RFC 4203

RFC 3473, RFC 3477

RFC 4204

according to GMPLS standards. A new RSVP-TE object LSP_TUNNEL_INTERFACE_ID was proposed to be used when signaling a new FA-LSP in the Path and Resv RSVP-TE messages [10]. The object contains two fields, that is, the FA-LSP identifier and the router ID.

In the event of a new λ-LSP to be set up, the head-end OCC must allocate an identifier for the interface associated to the yet to be created FA-LSP. Next, it originates an RSVP-TE Path message containing a LSP_TUNNEL_INTERFACE_ID object filled with the selected local interface identifier, along with the local optical node identifier. When the Path message arrives to the destination, the tail-end OCC must allocate an identifier for that FA-LSP end. This is called the remote FA-LSP interface identifier, which is reported back to the head-end within the RSVP-TE Resv message. As soon as the λ-LSP has been created, the head-end OCC advertises it as a forwarding adjacency by means of OSPF-TE. Being the FA-LSP bidirectional, it is also advertised by the tail-end OCC. All OCCs receiving the FA advertisement update its link state database adding a new link between the involved nodes. As an example, Fig. 3 depicts the signaling procedures between nodes A-F involving an existing FA-LSP between B-E. In this case, two new λ-LSPs (A-B, E-F) must be established to support the client LSP A-F, which will reuse part of the spare capacity in the FA-LSP B-E.

As mentioned before, λ-LSP’s length may be limited to maximize resource re-utilization in the network. To achieve such purposes, λ-LSP establishment must be dissociated from the client LSP setup procedure, allowing in this way the establishment of several underlying λ-LSPs while signaling only one client LSP request. To permit this separation between client and λ-LSP set-up, the head-end OCC evaluates the accumulated optical length, using loose hops in the ERO RSVP-TE object [8], deciding whether or not the whole route segment has to be divided into several underlying λ-LSPs. The same mechanism is also adopted when an intermediate λ-LSP should be created. Receiving an intermediate OCC an RSVP-TE Path message with the next hop set as loose, it must perform a route expansion computing the next route segment and decide if it should be divided into several underlying λ-LSPs.

In this work, both the CC and the RC CARISMA modules have been enhanced with the RSVP-TE extensions and the routing metric detailed above. After a new λ-LSP is established, OSPF-TE flooding updates the new state of those allocated wavelength-channels, also advertising the residual available bandwidth on the newly created FA-LSP. This FA-LSP will be available for route computation until the residual available bandwidth becomes zero or it does not support any client LSP, when the subjacent λ-LSP will automatically be torn down.

IV. EXPERIMENTAL RESULTS For the evaluation, two different transport network

topologies have been considered: a medium-sized

FA B‐E

PATH msg

λA-B λE-F

PATH msgIF_ID = A1

PATH msgPATH msg

IF_ID = E1

RESV msgIF_ID = F1

RESV msgRESV msg

RESV msgRESV msgIF_ID = B1

FA E‐F

FA A‐B

A B C D E F

Fig. 3 LSP signaling process involving FA-LSPs.

topology composed of 9 nodes and 11 links (Fig. 4, left), and a larger one with 16 nodes and 23 links (Fig. 4, right). In both scenarios, each link supports 8 bidirectional wavelengths.

Departing from these basic network scenarios, three different architectural solutions have been analyzed: all-optical, generic FA and opaque. The first solution stands for an all-optical network, where an end-to-end light-path with the whole wavelength capacity is established per client LSP request. The second solution, generic FA, identifies the GMPLS-controlled traffic grooming exactly as explained in the previous section. Finally, the third solution contemplates an opaque transport network, where only single-hop λ-LSPs are allowed. Once established, these LSPs are advertised as FA-LSPs permitting their re-use.

For the traffic characteristics, we consider that uniformly distributed client LSP requests arrive to the network following a Poisson process with mean Inter-Arrival Time (IAT) equal to 1/λ. Besides, connection duration follows an exponential distribution with mean Holding Time (HT) set to 1/µ. In particular, the requested BW of all incoming client LSP requests is considered to be 1/4 of the total wavelength capacity.

Fig. 5 illustrates the obtained C(H) function for the 9-Node network topology. The bar graph plots the probability that an incoming client LSP request has a certain number of hops, assuming availability of resources through the shortest path. As seen, there is a 40% probability that an incoming request traverses 2 hops from source to destination. In contrast, only 5% of the incoming requests would traverse 4 hops. This validates our assumptions where we stated that, by splitting very long FA-LSPs into shorter one’s resources are much more likely to be reused. Values greater than the network diameter (i.e., 4 hops in our scenario) have pH = 0.0 and are not depicted in the figure. To finally obtain CFA(H) we fix h = 0.5, as it provided the best network performance, while fulfilling our design criteria: CFA(2) + CFA(2) < CFA(4) and CFA(1) + CFA(2) < CFA(3).

For each architectural solution, Fig. 6 plots the obtained client LSP blocking probability in the 9-Node network topology as a function of the offered load to the

Fig. 4 9-No

Fig. 5 CFA(H) fu

Fig. 7 Blockingtopology.

EMULATETRANSPO

OUT‐OF‐

ode (left) and 16-N

unction in the 9-N

g probability vs.

OCOCC9

ED ORT PLANE

‐FIBER SCN

Node (right) trans

Node network top

offered load in

I‐NNI

2 OCC3

sport networks un

ology.

n the 16-Node ne

NMI‐A

100 Mbps Ethernet point‐to‐point link

nder evaluation.

Fig. 6topolog

etwork Fig. 8 O

OXC Emulator

EMUTRAN

OUT‐O

Blocking probagy.

O/E port usage vs

LATED NSPORT PLANE

OF‐FIBER SCN

ability vs. offere

s. offered load in

OCC8OCC9

ed load in the 9

the 16-Node netw

I‐NNI

NMI‐A

9-Node network

work topology.

OXC ulator

100 Mbps Ethernet point‐to‐point link

network (λ/µ = HT/IAT), normalized to a value of 200. As expected, significantly better resource usage is achieved when implementing FA-LSP capabilities in the network. For instance, if 0.5% client LSP blocking probability would have to be assured, a maximum load L=0.02 could be offered to a pure all-optical wavelength-routed optical network. Conversely, it could be further increased to L=0.5 when FA-LSPs are implemented, almost overlapping the opaque solution. This ∆L = 0.48 assesses experimentally the FA-LSP capabilities to automatically manage grooming actions in future transport networks, given the limitations of pure wavelength-routed optical networks to allocate incoming sub-lambda client LSP requests. In fact, as a whole wavelength is allocated in per 1/4 wavelength capacity client LSP request in the all-optical architecture, 3/4 of the total network capacity is directly wasted. Qualitatively speaking, this approximately results in four times less carried traffic by the network.

Fig. 7 shows the blocking probability results for each architectural solution, now in the 16-Node network topology. Moreover, CFA(H) has been particularized to let generic FA obtain the best performance. Again, the blocking probability figures are drastically reduced when implementing the FA-LSP functionality in the network. Here, fixing a blocking probability value around 0.5%, the offered load to the network in the generic FA and opaque solutions can be increased by 0.6 compared to the all-optical one.

As seen, generic FA and opaque solutions provide similar blocking probabilities in both 9-Node and 16-Node network scenarios. In order to also envisage the related network CAPEX in all-optical, generic FA and opaque solutions, Fig. 8 quantifies the E/O port usage in the 16-Node network topology. To this end, the Y axis has been normalized to the maximum number of E/O ports that could be equipped in the network (one per output wavelength at each node).

As seen, besides pleading for huge electronic routers, able to process all the incoming information at the nodes, the opaque solution rapidly requires a large number of E/O ports, since an optical bypass is not possible in the network. Conversely, although the all-optical solution leads to the lowest port usage, this is mandated by its poor bandwidth efficiency, which forces it to consume all available wavelength channels. Interestingly, the generic FA lies between both solutions. As seen before, it leads to similar PB than the opaque solution. Nonetheless, thanks to the optical bypass capability, it leads to a lower E/O port usage, thus reducing the total CAPEX. Note that for an offered load of 0.7, leading to the same PB around 0.5%, the E/O port usage is decreased by 15%.

V. CONCLUSIONS AND FUTURE WORK This paper reported the implementation and

experimental validation of the CARISMA test-bed, a GMPLS-controlled grooming-capable network. To start with, the paper reviewed the Forwarding Adjacency (FA)

entity, further elaborating on a FA creation cost function CFA(H) that dissociates the λ-LSP set up from the network signaling process, enhancing in this way the reuse of already established FA-LSPs.

For evaluation purposes, two alternative network scenarios have been configured, namely, a medium-sized 9-Node network and a larger one composed of 16 nodes. In particular, the 9-Node network has served to exemplify the rationale behind the dissociation between client and λ-LSP establishments according to the CFA(H) function. From the obtained experimental results, GMPLS-controlled multi-layer integrated routing drastically improves the network blocking probability compared to an all-optical network solution. Moreover, these benefits increase as the network topology gets larger, as identified when moving from the 9-Node to the 16-Node network. Compared to an opaque network solution the resulting network blocking probability remains almost equal. However, we have obtained that thanks to the optical bypass capability, GMPLS-controlled multi-layer integrated routing leads to lower E/O port usage, which is translated into lower network CAPEX.

The evaluation presented in this paper concerns single 9-Node and 16-Node ASON domains. Further work will extend the implementation of the FA-LSP functionality in larger multi-domain multi-layer network environments. It will be our goal to assess not only the performance, but also the scalability of the GMPLS-controlled grooming as the network gets larger, even spanning more than a single domain.

ACKNOWLEDGEMENTS The work presented in this paper has been partially

supported by the Spanish Science Ministry through the project ENGINE (TEC2008-02634).

REFERENCES [1] ITU-T Rec. G.8080/Y.1304, “Architecture for the

Automatically Switched Optical Networks”, Nov. 2001. [2] E. Mannie, “Generalized Multi-Protocol Label Switching

(GMPLS) Architecture”, IETF RFC 3945, Oct. 2004. [3] K. Zhu, B. Mukherjee, “Traffic grooming in an Optical

WDM Meshed Network”, IEEE Journal on Selected Areas in Communications, Jan. 2002.

[4] R. Dutta, G. Rouskas, “Traffic grooming in WDM networks: past and future”, IEEE Network, Nov. 2002.

[5] K. Kompella, Y. Rekhter, “Label Switched Paths (LSP) Hierarchy with Generalized Multi-Protocol Label Switching (GMPLS) Traffic Engineering (TE)”, IETF RFC 4206, Oct. 2005.

[6] J. Comellas. R. Martínez, J. Prat, V. Sales, G. Junyent, “Integrated IP/WDM routing in GMPLS-based optical networks”, IEEE Network, Mar. 2003.

[7] D. Katz, K. Kompella, D. Yeung, “Traffic Engineering (TE) Extensions to OSPF Version 2”, IETF RFC 3630, Sep. 2003.

[8] L. Berger, “Generalized Multi-Protocol Label Switching (GMPLS) Signaling Resource ReserVation Protocol-

Traffic Engineering (RSVP-TE) Extensions”, IETF RFC 3473, 2003.

[9] B. Rajagopalan, J. Luciani, D. Awduche, “IP over Optical Networks: A Framework”, IETF RFC 3717, Mar. 2004.

[10] K. Kompella, Y. Rekhter, “Signalling Unnumbered Links in Resource ReSerVation Protocol - Traffic Engineering (RSVP-TE)”, IETF RFC 3477, Jan. 2003.

[11] J. Perelló, E. Escalona, S. Spadaro, J. Comellas, G. Junyent, “Resource Discovery in ASON/GMPLS Transport Networks”, IEEE Communications Magazine, Aug. 2007.

[12] J. Lang, “Link Management Protocol (LMP)”, IETF RFC 4204, Oct. 2005.

[13] D. Harrington, R. Presuhn, and B. Wijnen, “An Architecture for Describing Simple Network Management Protocol (SNMP) Management Frameworks,” IETF RFC 3411, 2002.

[14] GNU Quagga Routing Software. http://www.quagga.net [15] K. Kompella, Y. Rekhter, “OSPF Extensions in Support of

Generalized Multi-Protocol Label Switching (GMPLS),” IETF RFC 4203, 2005.

Enhancing Performance of Optical Communication Systems with Advanced Optical

Signal Processing

Ivan Glesk University of Strathclyde/Electronic and Electrical Engineering Department, Glasgow, UK

Email: ivan.glesk@eee.strath.ac.uk

Marc Sorel and Anthony E. Kelly University of Glasgow/Electrical Engineering Department, Glasgow, UK

Email: {sorel, tkelly}@elec.gla.ac.uk

Paul R. Prucnal Princeton University/Electrical Engineering Department, Princeton, USA

Email: prucnal@princeton.edu

Abstract—Growing needs to transport large amount of data, penetration of multimedia into our daily lives, and quickly expanding e-Commerce sector triggered an unparallel demand for the new generation of fast, secure, and energy savvy communication networks. Today we already benefit from many advances which revolutionized data and voice communication. Commercially deployed Dense Wavelength Division Multiple Access (DWDMA) networks today are capable of transporting tens of Gigabits of data per second over a single WDM channel thus offering tremendous aggregate data throughputs over a single optical fibre. As a consequence, new bottlenecks have emerged at the fibre endpoints where data detection, routing, and switching must take place. Today’s routers use electronics to process all incoming optical traffic. However the available bandwidth offered by current electronics can no longer keep up with these rapidly growing demands. To address these challenges and with goal in mind to eliminate this bottleneck, the research community has been looking long and hard for appropriate alternative solutions. One of taken approaches can be described as optical signal processing. As we will demonstrate it can be very powerful tool to improve performance of advanced communication networks especially when coupled with technologies and approaches which will enable device integration and packaging. Index Terms—optical signal processing, optical CDMA, all-optical signal processing, optical XOR, device integration

I. INTRODUCTION

Optical code division multiple access (OCDMA) can provide variable data rates while delivering much higher channel count per number of used wavelengths than DWDM by avoiding problems of hard blocking. This paper describes two different applications of optical

signal processing which were developed to improve the performance of an incoherent Optical Code Division Multiple Access (OCDMA) system.

One application of an optical signal processing was to improve physical layer data security by implementing optical layer One-time Pad (OTP). OTP is an encryption algorithm where the plaintext is combined with a random key or a "pad" that is at exact same length as the plaintext and is used only once. If the key is truly random, never reused, and kept secret, OTP can be proven to be unbreakable [1]. A very promising approach of One-time Pad implementation is via optical XOR applied directly in the network physical layer. If is used in conjuction with OCDMA approach it can deliver increased data privacy, including channel isolation among the network users.

The second application of an optical signal processing - ultrafast all optical time gating, was used to improve system efficiency, maximize number of simultaneous users, and to improve signal to noise ratio in an OCDMA system.

The paper is organized the following way. Section II describes the implementation and experimental demonstration of the physical layer optical XOR and the performance demonstration of the OCDMA system with implemented OTP approach. In Section III is theoretically and experimentally shown how the OCDMA can benefit from optical signal conditioning by employing an ultrafast all-optical time gating. Section IV discusses our approaches and presents new results towards obtaining key integrated devices for the use in discussed approaches. Section V contains a brief summary.

II. DEMONSTRATION OF PHYSICAL LAYER OPTICAL ONE-TIME PAD

The reinforcement of security is one of the most urgent and critical issues in today networks. Currently, the

network security is offered in the application layer, and hence, is application-dependent. If the security can be enhanced in the network physical layer, it will become applications-independent, thus making networks more robust. For this reason, various approaches including optical cryptography have been gaining lot of attentions. One very promising alternative is OTP implemention via optical XOR directly in the network physical layer.

A. Hardware Implementation To demonstrate this, we first designed a “dual code”

transmitter/receiver pair, Tx/Rx-D (see Fig. 1(b)), in which one OCDMA code represents “1”s and a different code represents “0”s, having the capability to swap codes on a bit-to-bit bases during the data transmission. To

evaluate the performance, Tx/Rx-D were inserted into an OCDMA testbed (Fig. 1(a)) which uses (3,11) wavelength-hopping time-spreading prime codes of 73 ps chip size. The testbed has a modified network bus architecture. Additional Tx/Rx-A&B are used to introduce multi access interference, and a “designated” eavesdropper, Rx-E, to help conduct eavesdropping studies. Integrated part of the testbed is a picosecond multi-wavelength optical source; designed to generate three wavelengths by spectral slicing of 10 nm of supercontinuum with 200 GHz thin film filters (TFFs) cantered at 1550.12, 1551.72, and 1553.33 nm, and able to generate 5.1 ps FWHM optical pulses at a rate of OC-24 (~1.25 Gbit/s).

A “dual code” transmitter Tx-D consists of two Optical Encoders D1 & D2 (Fig. 1(b)), followed by two 10 Gb/s

LiNbO3 2x2 cross-bar switches. The first switch, driven by the RF data, generates optical data stream by “imprinting” RF bits “1”s and “0”s on two of (3,11) wavelength-hopping time-spreading family of OCDMA codes [2] d1(λ1 -1, λ2 -2, λ3 -3) and d2(λ1 -1, λ2 -3, λ3 -5), respectively. The second switch driven by the swapping key on the bit-to-bit bases selects which of the two OCDMA codes d1 or d2 will be representing and carrying bit “1”s and bit “0”s. It can be shown that the output of the Tx-D is a result of “XOR” operation between data bits and the swapping key bits as illustrated in Fig. 1(c). If the RF swapping key is randomly generated, transmission from Tx-D enhances channel isolation and data security/privacy in the optical layer by emulating the performance of the “One-time Pad” (OTP).

The swapping key we used was a 231-1 pseudorandom pattern (not truly random) which should be kept in mind when assessing the security aspects of the data transmission. An example of a scrambled data transmission between “dual code” Tx-D and Rx-D using optical layer XOR is shown in Fig. 1(c).

code d111code d201code d210code d100

Data XOR KeyRF KeyRF Data

Server

A “dual code” receiver Rx-D is shown in Fig. 1(b) and consists of two optical decoders D1 and D2; a 2x2 LiNbO3 switch is used to select and deliver the proper OCDMA code carrying optical data from the optical bus to the desired optical decoder D1 or D2 by means of applying a RF swapping key. To properly decode data, RF swapping key is identical to the one used by the transmitter Tx-D. The required linear input polarization was adjusted by a single PLC. No polarization mode dispersion was observed. If desired, polarization maintaining fiber for

RF Swapping Key 0011

OptEnc D2

RF Data: In 0101

OptDec D1d1d1d1d1

d2d2d2d2

RFData D

d2d1d2d1

d1d2d1d2

d1d2d2d1 • • d2d1

d1d2 • •

Switch Control“1” bar“0” cross

OptEnc D1

OptDec D2Opt

4 ×1 com

RF Code swapping

RF In Data B

RF In Data A

RF In Data D

ps multiw

avelength laser

2λ1 ,λ

2 , λ3

1% Taps

optical bus

Data APDOptDec A

Rx-AData B

PDOptDec BRx-B

Data EPDOptDec D1

Rx-E Eavesdrop Data D

RF Code swapping

OptDec D2

OptDec D1

Rx-D Secure

2psAOTG

Optical control

OptEnc D2

OptEnc D1

Tx-A: OptEnc A

Tx-B: OptEnc B

OCDMA representation

(λ1 - 1, λ2 - 2, λ3 - 3)

(λ1 - 1, λ2 - 3, λ3 - 5)(λ1 - 1, λ2 - 3, λ3 - 5)

Figure 1(a). Diagram of OCDMA testbed; (b) Example of a scrambled data transmission between Tx-D and Rx-D using optical layer XOR; (c) Resulted XOR output from the transmitter Tx-D.

optical bus implementation can be used instead of single mode fiber to avoid any polarization issues.

B. Experimental Results Figure 2(a) is an example of a bit error rate versus

received power obtained by Rx-D when Tx-D only (blue squares) or all other transmitters Tx-A & Tx-B (green

-26 -24 -22 -20 -18 -1610-14

Rx-D (231-1 swap),Tx-D onlyRx-D (231-1 swap), all Tx’sRx-E (no swap),Tx-D only

Receiver Power [dBm]

Figure 2(a). BER vs. received power for Rx-D & Rx-E.

circles), respectively, were present and RF swapping key of PRBS 231-1 was used during the data transmission. Bit error rate better than 10-12 was measured and no error floor was observed. To study performance of a novel Tx/Rx-D under different eavesdropping conditions, the receiver Rx-E (Fig. 1(a)) was designed to have its optical decoder identical with the optical decoder D1 also used by the receiver Rx-D. Depending on how the user D applies the RF key during his data transmission strongly affects the ability of the eavesdropper, with his receiver Rx-E, to eavesdrop on the user’s D transmission. Fig. 2(a) shows obtained BER for the receivers Rx-D and Rx-E under conditions as indicated in the inset of Fig. 2(a). When the pseudorandom swapping key used by the transmitter Tx-D was turned ON, the eavesdropper Rx-E was not be able to obtain any information, thus insuring secure data transmission between the Tx-D and Rx-D. However, if this swapping key was turned OFF, no XOR operation was performed by Tx-D, eavesdropper could eavesdrop on the user D transmission and in this case BER better than 10-12 was obtained by the eavesdropper (see red diamonds in Fig. 2(a)).

Figure 2(b) shows eye diagrams which were obtained by all receivers Rx-A, Rx-D, and Rx-E, respectively

Figure 2(b). Example of a bus traffic (top) and obtained eye diagrams for Rx-A, Rx-D and Rx-E.

during simultaneous data transmissions as well as a snapshot illustrating traffic on optical bus.

III. SIGNAL CONDITIONING BASED ON ULTRA FAST ALL-OPTICAL TIME GATING

To theoretically and experimentally demonstrate how signal conditioning via ultrafast all-optical time gating can improve the system efficiency, the number of simultaneous users, and the signal to noise ratio, we modified the described “dual code” OCDMA receiver Rx-D by amending it with a 2ps all-optical time gate. It is schematically shown in Fig. 3(a). The newly designed receiver operates as follows: Its conventional part based

Optical Control D

RFDataOut

Control

Decoded OCDMA DataMAI

Rx-D TOAD λ-filter

All-Optical Time Gate

1525 1545 1565 Wavelength [nm]

Det+Th

λ4 = 1544.06nmλ1 = 1550.12 λ2 = 1551.72λ3 = 1553.33

λ1 – λ3

Figure 3(a). Diagram of OCDMA receiver with all-optical time gate; Rx-D: conventional OCDMA receiver, D: optical tunable delay line;

Det+Th: photo detector with thresholder; SOA: semiconductor optical amplifier.

on the receiver Rx-D (see Fig. 3(a)) serves as a front end to decode incoming data from the network. The decoded signal consists of a correlation peak e.g., decoded OCDMA data send by transmitter Tx-D and the Multi Access Interference (MAI) noise from all other transmitting testbed users. To suppress MAI the decoded signal now enters the all-optical time gate (see Fig. 3). The time gate is based on a Terahertz Optical Asymmetric Demultiplexer (TOAD) [4] which is an ultrafast all-optical demultiplexer with its switching window SW (see Fig. 3 (c)) set approximately to ~2ps.

Time [ps]

Opticalcontrol

opticalmidpoint

ccw cw

OutMAI MAI

λ-filter (b) (c)

Figure 3(b). TOAD-based realization of 2ps all-optical time gate; (c) ~2ps time gating window; inset: TOAD input/output eye diagrams

indicating 6dB gain.

This was achieved by positioning a nonlinear element – a polarization insensitive semiconductor optical amplifier, SOA, at the distance Δx from the Sagnac loop midpoint. By biasing the SOA with current of I = 200mA the time

gate delivered 6dB of input-output signal gain. To properly time overlap the TOAD switching window SW, (sees Fig 3(c)) with the correlation peak e.g., the decoded OCDMA data (see Fig 3(b) for illustration), the optical control pulse was delayed by an optical tunable delay line D (see Fig. 3(a)).

Under these conditions only data from the transmitter Tx-D are demultiplexed and after passing the optical filter will appear at the TOAD output port with the gain of 6dB. The filter serves as an optical high pass for λ1-λ3 while blocking the optical control at λ4 (see Fig. 3(a)). After passing the filter, the “clean” OCDMA data, now without MAI, are photodetected, thresholded and then fed into a bit error tester for further performance measurements. To match the code chip rate, a 12.5 GHz optical receiver was selected as a frontend photo detector for the 12.5 Gbit/s BER tester.

Fig. 4(a) shows the output of a conventional OCDMA receiver Rx-D not equipped with the 2ps time gate

Figure 4(a). Eye diagram obtained by conventional receiver Rx-D.

showing an eye diagram when Rx-D is tuned to listen to the transmitter Tx-D. The Rx-D output was observed using bandwidth limited sampling oscilloscope. The peaks in Fig. 4(a) indicate the autocorrelation, e.g., the decoded OCDMA data. The other pulses in Fig. 4(a) are cross correlations from the remaining transmitters broadcasting on the network and represent the unwanted MAI noise. Fig. 4(b) shows the elimination of the MAI

Figure 4(b). Eye diagram obtained by receiver Rx-D equipped with 2ps time gate.

noise when the OCDMA receiver equipped with 2ps time gate was used. Here, only the autocorrelation peaks, e.g. the decoded OCDMA data received from the transmitter Tx-D are seen, no MAI nose is present.

The performance calculations for a conventional OCDMA receiver (without an ultrafast all-optical signal processing) indicate (see Fig. 5 curve-ii) an error-free operation of the testbed for up to four simultaneous users. Using PRBS 231-1 we measured BER better than 10-12, no

Conventional OCDMA receiver(ii)

OCDMA receiver with 2ps time gate

0 10 20 30

Number of simultaneous users

Figure 5. Calculated probability of error as function of the number of simultaneous users for receiver Rx-D: (i) with and (ii) without

2ps time gate.

error floor was observed (see Fig. 6). The simulations also predict that the error rate will increase with an increasing number of simultaneous users thus limiting the

Received Power [dBm]

-28 -26 -24 -22 -20 -18 -16

Rx-D (231 -1); all TxRx-D (231 -1); all Tx + 2ps sampling

Figure 6. Measured BER with all simultaneous users present: circles - conventional Rx-D; diamonds - Rx-D + 2ps all-optical

time gate.

network scalability and the maximum number of simultaneous users. However, calculations also predict that this limitation can be dramatically reduced if MAI noise is eliminated. This was demonstrated with OCDMA receiver equipped with 2ps time gate. In this case calculation results shown in Fig. 5 curve-(i) indicate a significant increase in number of simultaneous testbed user as well as a substantial improvement in the BER system performance. Example: for a given BER, say 10-9, the number of simultaneous testbed users will more than triple and should grow from tree to ten, all without sacrificing the BER performance. This was experimentally confirmed by BER measurements which are shown in Fig. 6.

IV. MONOLITHIC INTEGRATION OF KEY FUNCTIONALITIES

To make described approaches more easily applicable, practical and inexpensive a monolithic integration of

proposed devices is desired and necessary. We are currently working on integration of a cascaded Mach-Zehnder structure for the One-time Pad application as well as on a TOAD-like ultrafast all-optical sampling gate with a tunable wavelength filter at its output. This wavelength filter will ensure that the gate control signal can be filtered out and will not interfere with the correlation peak at the gating device output. The filter tunability will also offer more flexibility during various device applications. Our fabrication is based on a monolithically integrated

III-V semiconductor platform. This technological approach enables the fabrication of devices with high performance, compactness and low-cost. In particular, the bandgap of quantum well gain media can be easily engineered to enhance design flexibility and integration capabilities.

Devices requiring integration of multiple bandgaps on a single chip can be fabricated using a variety of integration technologies such as selective area regrowth [7], twin-waveguide structure [8], and butt-coupling scheme [9]. Here we use a quantum-well intermixing (QWI) technique, which allows postgrowth modification of the bandgap in quantum confined heterostructures via a compositional disordering process between the QWs and barriers [10].

Fig. 6 reports the schematics of a monolithically integrated all-optical sampling gate configured here as a Mach-Zehnder Interferometer (MZI). The device was

Figure 6. Schematic of a MZI device (left) and output powers of the bar and cross states as a function of the injected current on one of the phase

shifters (right).

fabricated in the InGaAs–Al-GaInAs material system, which exhibits a larger conduction band offset, a larger differential gain, and better intrinsic thermal characteristics [11] compared to its phosphorus-based counterpart. The access waveguides were quantum-well intermixed to 1450 nm in order to minimize the absorption losses, while the bandgap of the two phase shifters on the MZI arms was kept at 1550 nm. Under

forward biasing, the phase shifters exhibit high phase shift efficiency of 140° /mA /mm and the MZI switches from the cross to the bar state with an extinction ratio of 20 dB and total on-chip insertion losses of < 10 dB [12]. When the phase shifters are operated under reversed voltage - i.e., QCSE [13] – the MZI exhibits switching performance up to 20 GHz and beyond therefore the device switching speed is sufficient for the proposed One-Time Pad operation. (Please remember, the encryption is done on the data e.g., OCDMA code rate, not the chip rate.

Another very useful device in this context is an Integrated Asymmetric Mach-Zehnder Interferometer (IAMZI) multiplexer/demultiplexer shown in Fig. 7 [8]. It consists of two multimode interferometer couplers and 500 um-long phase shifter elements placed along the interferometer arms. The upper arm is a straight waveguide, while the lower arm is a curved waveguide with radius of 600μm, providing a total path difference of 455μm, which corresponds to a channel spacing of 100GHz (0.8nm). Injecting current into the interferometric arm(s) cause(s) a change in the refractive index of the waveguide and therefore enables the device to work as a tunable (de)multiplexer [9].

Figure 7. Optical picture of an asymmetric MZI (left) and output powers of the bar and cross states as a function of the input wavelength (right).

In our demonstrated application the IAMZI will be

used at the output of the MZI-based gate to filter out (DeMUX) the MZI-gate control signal (compare with the originally implemented approach in Fig. 3(a)).

Ultimately at the end of this effort, both devices will be monolithically integrated together on a single chip and then packaged before their use in all follow up experiments.

V. CONCLUSIONS

Undoubtedly, features like security, speed, and energy efficiency are the underpinning “must have” properties of any future communication networks which will be used as the supporting infrastructure for the upcoming era of Digital Economy.

In this paper we described two experimentally demonstrated applications of advanced optical signal processing techniques which were developed to improve system properties, its security, and the overall performance.

One example of such application – the introduction of the optical XOR gate was used to demonstrate increased data security of the optical network by way of implementing One-time Pad functionality in the network transport layer.

The second example – an ultrafast all optical signal conditioning - was introduced to demonstrate improved performance of the incoherent optical CDMA system namely its scalability, including number of simultaneous users, which resulted in the improved power budget via enhanced signal to noise ratio.

In the last section we reported our recent results in the area of a device fabrication. We also discussed our ongoing effort towards monolithic integration of the demonstrated ultrafast all-optical multifunctional devices which we believe when successful would lead to ease the applicability of demonstrated optical approaches, will make them more practical, and also less expensive to implement.

REFERENCES

[1] C. E. Shannon: “Communication theory of secrecy system,” Bell Sys. Tech. J., vol. 28, pp. 656-715, 1949.

[2] G.-C. Yang, W. C. Kwong: Prime Codes with applications to CDMA Optical and Wireless Networks, Artech House, Norwood, MA, 2002.

[3] Y.-K. Huang, “Integrated holographic encoder for wavelength-hopping time-spreading optical CDMA,” Photon. Technol. Lett., vol. 17, pp. 825-827, 2005.

[4] J. P. Sokoloff, P. R. Prucnal, I. Glesk, and M. Kane, “A Terahertz Optical Asymmetric Demultiplexer (TOAD),” Photon. Technol. Lett., vol. 5, pp. 787-790, 1993.

[5] I. Glesk, Y.-K. Huang, C.-S. Brès, P.R. Prucnal, “Design and demonstration of a novel Optical CDMA platform for avionics applications,” Opt. Com., vol. 271, pp. 65-70, 2007.

[6] I. Glesk, C.-S. Brès, Y.-K. Huang, D. Rand, and P. R. Prucnal “Demonstration of a 1.25 Gbps Incoherent 2D-OCDMA Communication Platform with Bit-To-Bit Code Swapping,” CLEO/IQEC and PhAST Technical Digest on CDROM (The Optical Society of America, Washington, DC, 2006), CWH2.

[7] T. Tanbun-Ek, P. F. Sciortino, A. M. Sergent, K.W.Wecht, P.Wisk, Y. K. Chen, C. G. Bethea, and S. K. Sputz, “DFB lasers integrated with Mach-Zehnder optical modulator fabricated by selective area growth MOVPE technique,” IEEE Photon. Technol. Lett., vol. 7, pp. 1019–1021, 1995.

[8] P. V. Studenkov, M. R. Gokhale, W. Lin, I. Glesk, P. R. Prucnal, and S. R. Forrest, “Monolithic integration of an all-optical Mach-Zehnder demultiplexer using an

asymmetric twin-waveguide structure,” IEEE Photon. Technol. Lett., vol. 13, pp. 600–602, 2001.

[9] E. Jahn, N. Agrawal, M. Arbert, H. J. Ehrke, D. Franke, R. Ludwig, W. Pieper, H. G. Weber, and C. M. Weinert, “40 Gbit/s all-optical demultiplexing using a monolithically integrated Mach-Zehnder interferometer with semiconductor amplifiers,” Electron. Lett., vol. 31, pp. 1857–1858, 1995.

[10] W. D. Laidig, N. Holonyak Jr., M. D. Camras, K. Hess, J. J. Coleman, P. D. Dapkus, and J. Bardeen, “Disorder of an AlAs-GaAs superlattice by impurity diffusion,” Appl. Phys. Lett., vol. 38, pp. 776–778, 1981.

[11] B. Borchert, R. Gessner, and B. Stegmuller, “Advanced 1.55 μm quantum-well GaInAlAs laser diodes with enchanced performance,” Jpn. J. Appl. Phys., vol. 33, pp. 1034–1039, 1994.

[12] H. Y. Wong, W. K. Tan, A.C. Bryce, J. H. Marsh, J. M. Arnold, and M. Sorel, “Monolithically Integrated InGaAs/AlGaInAs Mach-Zehnder Interferometer Optical Switch Using Quantum Well Intermixing,” IEEE Photonics Technology Letters, vol. 17, pp. 999-1002, 2005.

[13] D. A. B. Miller, J. S. Weiner, and D. S. Chelma, “Electric-Field dependence of linear optical properties in quantum well structures: Waveguide electroabsorption and sum rules,” IEEE J. Quantum Electron., vol. QE-22, pp. 1816–1829, 1986.

[14] H. Y. Wong, W. K. Tan, A.C. Bryce, J. H. Marsh, J. M. Arnold, J. M. Krysa and M. Sorel, “Current Injection Tunable Monolithically Integrated InGaAs/InAlGaAs Asymmetric Mach-Zehnder Interferometer using Quantum Well Intermixing,” IEEE Photonics Technology Letters, vol. 17, pp. 1256-1260, 2005.

[15] S. Matsuo, Y. Yoshikuni, T. Segawa, Y. Ohiso, and H. Okamoto, “A widely tunable optical filter using ladder-type structure,” IEEE Photon. Technol. Lett., vol. 15, pp. 1114–1116, 2003.

Ivan Glesk received the B.Sc., M.Sc. and Ph.D. degree in Quantum Electronics and Optics from Comenius University in Bratislava and D.Sc. degree from the Slovak Academy of Sciences 1998.

In 1986, he joined Comenius University and later became Professor of Physics at the Department of Experimental Physics where he conducted his research in the areas of nonlinear optics, laser physics, and LIDAR sensing of the Atmosphere. As a Recipient of the IREX Fellowship he was a Visiting Fellow at the Department of Mechanical and Aerospace Engineering at Princeton University in 1990-1991. After that he joined the Department of Electrical Engineering at Princeton University where he was Senior Research Scholar and Manager of the Lightwave Communication Research Laboratory. In 2007 he joined Electronic and Electrical Engineering Department, University of Strathclyde, Glasgow. UK as a Professor of Broadband Communication Systems. His current research interests encompass broadband communication systems, optical interconnects; and all-optical switching and its applications.

Prof Glesk has co-authored 16 book chapters, over 230 scientific publications, presented 31 invited talks and lectures at various international conferences and institutions and holds 5 patents. He is IEEE Senior Member and member of several international committees. Since 1998, he has been a Chairman of Slovak Commission for Optics.

Marc Sorel graduated cum laude in 1995 in Electronic Engineering and received the Ph.D. degree in Electronic Engineering and Computer Science in 1999 from the Università di Pavia, Italy.

In 1998 he joined the Optoelectronics Group at the University of Glasgow and he was awarded a Marie-Curie fellowship in 1999 for research on integrated optical gyroscopes. He currently holds a senior lecturer position at the Electrical Engineering Department University of Glasgow, UK. His research activity includes fast dynamics in semiconductor lasers, ring resonators, integrated semiconductor device fabrication and characterization. He is principal investigator for the Glasgow contribution to the FP6 IOLOS and SPLASH projects, as well as several national projects on integrated semiconductor devices - with the demonstration of ultrafast all-optical switching and laser dynamics being a particular highlight.

Dr. Sorel has published over 50 papers in peer-reviewed journals and over 100 conference papers.

Anthony E. Kelly received the B.Sc., M.Sc. and Ph.D. degrees from the University of Strathclyde.

He previously worked at British Telecom Laboratories and Corning and was a cofounder of Kamelian Ltd and Amphotonix Ltd., Glasgow, UK. He currently holds a senior lecturer position at the Electrical Engineering Department University of Glasgow, UK. His current research is in the use of

semiconductor optical amplifiers for PONs, optical burst switching, and ultrafast optical switching.

Dr. Kelly has published over 100 journal and conference papers on a range of optoelectronic devices and systems and holds a number of patents.

Paul R. Prucnal received the A.B. degree from Bowdoin College, Brunswick, ME, and the M.S., M.Phil., and Ph.D. degrees from Columbia University, New York.

He was a faculty member until 1988 at Columbia University. Then he joined Princeton University, Princeton, NJ, USA as a Professor of electrical engineering, where he was the Founding Director of Princeton’s Center for Photonics and Optoelectronic Materials from 1990 to 1992. He was also a Visiting Professor at the University of Tokyo and University of Parma. His research interests include optical code-division multiple access as well as physical layer security in optical networks.

Prof. Prucnal is a Fellow of the Optical Society of America (OSA) and Fellow of IEEE. He is also an Area Editor of the IEEE Transactions on Communications for Optical Networks. He was a recipient of the Rudolf Kingslake Medal from the Society of Photonic and Instrumentation Engineers. He is the author or coauthor of more than 200 journal papers and holds 17 patents.

Abstract— The capacity growth driven by increasing traffic

and the introduction of new devices and technologies has resulted in new functionalities in optical networks. The property of transparency enables compatibility between different system generations and the coexistence of multiple bit-rates without raising the global network cost. In a transparent network, the signals travel through many links and nodes without the need for opto-electronic regeneration, accumulating physical impairments. Therefore, standard point-to-point test-bed set-ups are no longer adequate to emulate signals propagating in transparent optical networks. New test-beds integrating the different heterogeneity cases due to the use of transparent cross-connects becomes necessary. In this paper, we present various test-bed realizations accounting for the following network heterogeneities: fiber link heterogeneity and neighbor changes while a signal propagates in the fiber. To be able to emulate such network heterogeneity, we present a double-loop set-up. Such a set-up is then used to assess the propagation results obtained by simulation and validate the quality of transmission (QoT) estimators.

Index Terms— Optical networks, Quality of Transmission (QoT) estimators.

I. INTRODUCTION

he improvement of transmission devices enables bothsignal propagation over longer distances and also increases in the transported bit-rate. Improvements on

optical switches for which the transparent pass-through and wavelength reconfiguration of optical signals is possible, have enabled greater flexibility and automation. Key characteristics of future networks include: higher transported capacity, greater ease of network upgradeability and greater flexibility to different technology evolution paths (i.e., transmission and/or modulation format). Underpinning all of these attributes is the advantage wavelength division multiplexing (WDM) technology which allows for the propagation of many channels on the same medium using shared resources. These properties reduce the number of devices that need to be changed to accommodate network evolution and hence the total cost over

the system lifetime. Today transparent networks are often used to cope with these requirements. Transparency results in reduced opto-electronic (OEO) conversion. As OEO devices are sensitive to signal formats, their reduction guarantees greater network flexibility.

Another trend, leading to a reduction in maintenance costs, is automatic reconfiguration of the network in response to the arrival of new connections or to the path recovery after a failure. But, to ensure the automatic reconfiguration of a transparent network, the control plane (which must exchange the required information to manage the whole network) needs the knowledge of the different impairments acting on the signal to set-up. Quality of transmission (QoT) estimators have been introduced as tools that may be used to evaluate the signal state at reception after propagation in the network. Recently, various estimators have been proposed in the literature, including dependence on the bit-rate and the modulation format of the transmitted signal, and also on the physical characteristics of the traversed systems. These estimators can be obtained using laboratory experiments [1] or simulation results, [2], [3]. A key challenge for laboratory methods is to achieve a wide range of network scenarios, which on the other hand is readily done by simulation. At the same time, the use of only simulations to validate various estimators is not fulfilling because the simulation model can be incomplete or some discrepancy between simulations and physical transmission can appear. Indeed an optical network spreading over several thousand kilometers can be acombination of different generations of systems having different characteristics (i.e. amplifier gain ripple, in-line fiber loss and/or polarization mode dispersion, PMD). Hence usual point-to-point test-bed transmission experiments might not be representative of a typical transparent transmission channel in a mesh network. To validate a QoT estimator by the use of an experimental set-up, one may need to consider the heterogeneity found in deployed systems and the associated transported signals. In this paper we describe the realization of diverse experimental test-beds to assess a QoT estimator obtained by

Advanced Test-beds to Validate Physical Estimators in Heterogeneous Long Haul

Transparent Optical Networks

Annalisa Morea, Florence Leplingard, Jean-Christophe Antona, Pascal Henri, Thierry Zami

Alcatel-Lucent Bell Labs France, Route de Villejust, 91620 Nozay FRANCE e-mail: Annalisa.Morea@Alcatel-Lucent.com

Daniel C. Kilper

Alcatel-Lucent Bell Labs USA, 791 Holmdel-Keyport RD, Holmdel, NJ 07733 UNITED STATES e-mail: Dan.Kilper@Alcatel-Lucent.com

simulations. Discrepancy between simulation and experimental results has been used to refine the estimator. The paper is organized as follows: Section II describes the establishment of a quality of transmission estimator by the means of physical measurements and simulations of signal propagation; Section III describes the experimental set-up for a network with fiber heterogeneity and explains how the quality estimator is calibrated thanks to the experimental measurements; Section IV describes physical set-ups taking into account the impact of routing on the transmission performance, again the agreement between simulations and experimental measurements are demonstrated. Conclusions are then provided in Section V.

II. SECTION 2: QUALITY OF TRANSMISSION ESTIMATOR

A routing algorithm working in a transparent network has to search a transparent path (called ‘lightpath’ in the following of the paper) connecting two nodes in the network. If the required path is not feasible as an end-to-end transparent channel (either for wavelength or for physical feasibility constraints), the connection can be split in more than one lightpath by the means of OEO devices [4].

QoT estimators are integrated into the routing algorithms, which can be run off-line during the network planning phase or on-line by the control plane operating dynamically. Thus, if the QoT is used off-line, the estimator can be computationally complex to compute the signal state at detection (for example based on the non-linear Shrödinger equation [5]); however, when it is used by the control plane, the QoT estimator has more constraining requirements: computational simplicity and accurate physical impairment assessment.

Reliability is critical for a QoT estimator since it will dictate the lightpath configuration for the entire network. Indeed, in the case of underestimation (i.e. a connection is estimated as unfeasible when in reality it is feasible), more regenerators than needed would be installed and it may decrease the cost-efficiency of the whole network [2]. In the case ofoverestimation (i.e. a connection is estimated as feasible when in reality it is unfeasible), fewer than required regenerators are planned and several connections may have to be lit in order to find a working channel or the connections may be faulty.

The estimator presented in this paper is conceived to be integrated into the control plane and for this reason it is expressed by a simple analytic formula using cumulative parameters describing major physical impairments. At the same time, a precision value is associated to the estimator to determine its accuracy.

A. Estimator for 10 Gb/s NRZ modulation format

Here we present the method to establish the estimator developed for non-return to zero (NRZ) modulated signals at 10 Gb/s on a standard single mode fiber (SMF) [6]. This method is adapted to any kind of heterogeneity (i.e. modulation format, bit-rate, fiber type…) as shown in the next sections.

The physical performance of a signal is usually measured

through the Bit-Error Rate (BER), which is the ratio of erroneously transmitted bits over the total number of transmitted bits. This BER is often expressed in terms of the Q-factor, using the conversion relationship given in Equation (1) [7]:

⎟⎠

⎞⎜⎝

⎛=22

1 QerfcBER (1)

where Erfc(⋅) is the complementary error function. The Q-factor itself is generally expressed in a decibel scale, with Q2

dB= 20Log10(Q). Hence, a Q-factor of 12.6dB corresponds to a BER of 10-5. Note that (1) is only exact when the noise is Gaussian, which is not generally the case in optical systems. However, the Gaussian approximation is sufficiently accurate in most cases. This expression in decibels is often used to maintain consistency with the linear noise accumulation model, i.e.: in good approximation, the Q-factor scales linearly with the optical to noise ratio (OSNR), when also expressed in decibels [8] In the following, the proposed QoT estimator is aimed at predicting such a Q-factor in decibel scale.

In a transparent network, to determine whether a transparent connection is feasible, the Q-factor after transmission is usually taken to be higher than a given threshold. This ensures correct reception; which otherwise would require optoelectronic regeneration at some point within the lightpath.

The Q-factor depends on impairments experienced by the signal during the propagation, including: amplifiedspontaneous emission (ASE) noise, accumulated chromatic dispersion (Dres), non-linear signal distortion (expressed using the non-linear phase, φnl), PMD, and cross-talk (Xtalk). The Q-factor can be written using Equation (2), for more details see [4]:

)),,,(

),,,(),(

XtalkPMDDOSNR

OSNRDQ

XtalkPMDDgOSNRDQ

nlresQref

nlresref

nlresnlres

ϕζϕϕζ−⋅+

=+= (2)

where OSNR is the Optical Signal to Noise Ratio after transmission, assuming a noise power level calculated within a 0.1nm bandwidth, OSNRreq

Qref is the required level of OSNR at the end of a given lightpath to get a reference value of the Q-factor (Qref, hereafter Qref=12.6 dB for a BER of 10-5) and depends on the accumulated impairments. Let OSNRreq,BtB

be the OSNR ensuring a Q-factor equal to Qref in the back-to-back configuration (measured using a transmitter and receiver connected together without transmission); let PenQref represent the OSNR penalty due to the transmission, representing the required extra OSNR to obtain Q=Qref [8]. So, we obtain Equation 3:

)),,,(

(),( ,

XtalkPMDDPen

OSNROSNRDQQ

nlresQref

QrefBtBreqnlresref

ϕϕζ

−⋅+= (3)

The penalties due to noise, PMD, Xtalk and filtering can be

taken to be independent, as a first approximation; while distortions due to Dres and φnl depend on their coupled values. Thus Equation 3 becomes:

))()(),(

(),( ,

XtalkPenPMDPenDPen

OSNROSNRDQQ

QrefQrefnlresQref

QrefBtBreqnlresref

−−−

−⋅+=

ϕϕζ (4)

Methods to compute penalties associated with PMD and Xtalk are already presented in the literature [9], [10]; in [4] we explained how these penalties are added. Here we focus our analysis on Pen(Dres, φnl), where Dres is the accumulated dispersion on transmission and compensation fibers and φnl is obtained using [11], obtaining a quadratic function as a function of Dres. Then a, b, c coefficients are fitted by a second (or forth) order polynomial as a function of φnl. The estimator precision (and its complexity) is determined by the polynomial degree. The final function is given by Equation 5:

[ ]∑ ∑= ==

resnlresnlresnlresnlQref

DaDaDaDPen

φ φφφ (5)

with 2 ≤ k ≤ 4.

1) Simulation set-up To define a QoT estimator, often physical performance

results, obtained by the means of a simulator, are collected and the relationship between physical effects and the Q-factor is established.

The use of numerical simulation to define a QoT estimator can incorporate the emulation of a wide range of network scenarios. A network may be composed of links having a variable number of spans, each of which with a proper in-line dispersion-compensation. By using a different simulation configurations, it becomes possible to study various configurations in the number of spans per link and span length, and so the impact of each physical parameter such as Dres, φnl, OSNR, and their possible variations.

The simulation set-up used here is based on a 10.7 Gb/s SMF-based WDM transmission system emulated by a recirculating loop (Fig. 1). Before entering in the loop, a dispersion pre-compensation module (‘Pre-comp’ in Fig. 1) presents an accumulated dispersion of -870 ps/nm at1550.12nm. The loop is made of three 96 km SMF spans. Amplification stages following each fiber span include double stage amplifiers and dispersion compensation fiber (DCF1 in Fig. 1) to compensate respectively for in-line losses and accumulated dispersion (maintaining a residual dispersion per span of +100 ps/nm for the central channel at 1550.12 nm). At the end of the loop section, another DCF (DCF2 in Fig. 1) ensures a zero 0 ps/nm accumulated dispersion per loop section for the central channel. Finally, a chromatic dispersion post-compensation module (‘post-comp’ in Fig. 1) isintroduced before the receiver to vary the total accumulated dispersion Dres over the transmission lightpath. 21 channels propagate in the fiber, spaced by 50 GHz. Each channel is non-return-to-zero (NRZ) modulated at 10.7 Gb/s with a De

Bruijn sequence of 64 bits and an extinction ratio of 13 dB. Various time delays are considered between WDM channels randomly chosen between 0 and 6400 ps at the ingress node to emulate independent generation of information between channels (one simulation per each time delay).

The dispersion map uncertainties are emulated considering a deviation from the nominal value for each dispersion map. obtained by changing the nominal values of SMF lengths (the delta follows a Gaussian distribution with standard deviation of 5 km) and the nominal value of fiber dispersion (the variation follows a Gaussian distribution with standard deviation of 0.2 and 0.15 ps/nm/km for SMF and DCF1/2, respectively). The nominal channel power is 0 and 3 dBm in SMF fibers, while it is decreased by 7 dB in DCF. Amplifiers have flat gain over the C-band and noise is generated at the receiver side. Because of the large value of chromatic dispersion of SMF, four wave mixing (FWM) impact is weak compared to self- and cross- phase modulation (SPM and XPM, respectively) and it is not emulated in the simulation model. This set-up considers the worst case for thetransmission performance of lightpaths in a physical network as it considers a perfect in-line residual chromatic dispersion reset to zero and all channels propagate together along the paths (worst case of cross phase modulation, XPM, penalty). For a given distance and channel power, 800 simulations are performed using a specific dispersion map and chromatic dispersion uncertainties for each simulation. The OSNR penalty was determined by interpolating the estimation of penalties over the OSNR due to Dres and φnl. Fig. 2 shows the difference between calculated and simulated penalties. The errors between penalties are within ±0.5 dB for 84.4% of lightpaths.

After having simulated different scenarios to establish the QoT-estimator, it is important to validate it by the means of an experimental set-up. In fact, discrepancies between simulation results and measurements are likely to occur because of the simplified models chosen to emulate transmitters, receivers, fiber features or physical effects, and of the partial knowledge and heterogeneity of practical devices. It is thus important to dimension the impact of such discrepancies to be able to reliably extrapolate simulation results to transmission in the field and build accurate performance estimators.

100 km21 wavelengths

Pre-compPost-comp

Polarization

Scrambler

100 21 wavelengths

50GHz spacing

Pre-compPost-comp

A.O.A.O.

Polarization

Scrambler

A.O.A.O.

100 km21 wavelengths

Pre-compPost-comp

A.O.A.O.

Polarization

Scrambler

A.O.A.O.

100 21 wavelengths

50GHz spacing

Pre-compPost-comp

A.O.A.O.

Polarization

Scrambler

A.O.A.O.

Figure 1: Simulation and experimental set-up of the loop used to emulate the signal propagation in an optical network having in-line chromatic dispersion compensation and only SMF transmission fiber.

-0.3 0.

∆ Pen = Pen(simulated)-Pen(calculated) (dB)

-0.3 0.

∆ Pen = Pen(simulated)-Pen(calculated) (dB)∆ Pen = Pen(simulated)-Pen(calculated) (dB)

Figure 2: Error on the OSNR penalties obtained comparing simulated penalties (given using numerical propagation) and calculated (given using QoT estimator). The error follows a Gaussian distribution and penalties errors are within ±0.5 dB for 84.4% of lightpahts.

2) Experimental set-up The experimental set-up is obtained using the same set-up

shown in Fig. 1. But this time, each loop is made of three 100 km long spans of SMF fiber.

Each channel is NRZ modulated with a pseudo-random binary sequence (PRBS) of 223-1 bits at 10.7 Gb/s, with nominal power of 4 dBm in SMF fibers and -3 dBm in DCF. The signal performance is measured after 1 to 8 loops, corresponding to 300 to 2400 km with a step of 300 km. A polarization scrambler is placed into the loop to reduce loop polarization effects. Again, as was done for the estimator obtained with simulation results, measuring the OSNR penalty we obtain by interpolation the estimation of penalties over the OSNR due to Dres and φnl. In Fig. 3 we plot the difference between the measured and estimated OSNR penalties for the same values of (Dres and φnl); we observe an error following a Gaussian trend and the error is within ±0.5 dB for 91.5% of samples, ±1 dB for 98.6% and ±2 dB for 99.97%. Compared with the errors of the estimator obtained by simulation, the error distribution is narrower because this experimental set-up does not emulate the same variety of configurations as numerical simulations do and the interpolation of a limited heterogeneity of accumulated physical parameters is easier to interpolate.

∆∆∆∆Pen = Pen(measured)- Pen(calculated) (dB)

Figure 3: Error on the OSNR penalties obtained comparing measured penalties and calculated (given using QoT estimator). The error follows a Gaussian distribution and penalties errors are within ±0.5 dB for 91.5% of lightpaths.

0 0.5 1 1.5 2

Nonlinear phase (rad)

0 0.5 1 1.5 2

Figure 4: A comparison of the OSNR penalties obtained with experimental measurements (triangles) and simulations (squares) is depicted for 0 ps/nm Dres

at 1550.12 nm and φnl varying from 0 to 1.5 rad.

3) Simulated compared to experimental results To compare the two estimators obtained by simulation and by

experimental set-up, we correlate both results simulating the circulating loop using the same dispersion map as was used in the experimental set-up. In Fig. 4 the comparison of the OSNR penalties due to Dres and φnl for experimental measurements (triangles) and simulations (squares) is depicted for 0 ps/nm Dres at 1550.12 nm and φnl varying from 0 to 1.5 rad. From Fig. 4, the good agreement between the two estimators indicates a similar accord between simulation and experimental set-ups for other network configurations in the presence of SMF fiber types; hence the proposed estimator might be reliably used by a control plane to automatically compute a feasible lightpath in SMF based transparent networks.

III. DIVERSITY OF FIBER TYPE

Many optical networks are made of different fiber types with different characteristics, because the whole network is not built at the same time and during upgrades better performing fibers than older generations can become available.

If the diversity includes only the fiber generation (for instance SMF of old and new generations differing mainly in in-line attenuation and differential group delay, DGD) the previously developed estimator is able to account for this diversity because the impact of non-linear effects is similar. A more complicated task is to account for different fiber types in a network, for example the presence of SMF and Large Effective Area Fiber (LEAF) with quite different values of chromatic dispersion around 1530nm (from 16 to 2.5ps/nm/km) [12]. A good performance estimator has to consider the difference of link engineering for both fiber types and, then, the different impact of non-linear effects. LEAF fiber, indeed, is very sensitive to inter-channel nonlinear effects as compared to SMF, because of its low chromatic dispersion value around the transmission comb. For this reason a signal propagating through a few spans of LEAF can degrade significantly compared to the same distance traveled in only SMF spans. Fig. 5 shows the experimental measurements of OSNR penalties for a BER of 10-5 as a function of φnl after propagation in pure SMF-links and pure LEAF-links, for 10G NRZ channels separated by 50GHz.

0 0.50 1.00 1.50

Non linear phase, ϕϕϕϕnl (rad)

LEAF @ 0 dBm/ch @ 3 dBm/chSMF @ 0 dBm/ch @ 3 dBm/ch

0 0.50 1.00 1.50

LEAF @ 0 dBm/ch @ 3 dBm/chSMF @ 0 dBm/ch @ 3 dBm/ch

Figure 5: Example of different OSNR penalties due to non-linear effects after propagation in in pure SMF-links and pure LEAF-links. Results are obtained by experimental measures of OSNR penalties as a function of non-linear phase (φnl).

In the following, we explain the simulation and experimental set-ups used to establish and validate a QoT estimator adapted for transparent networks having fiber type mixing. A key feature in the following studies is the presence of only one kind of fiber per link (section between two adjacent nodes).

Today, in the literature [13] and [14] present theoretical studies and experimental assessment of impairments due to non-linear effects in heterogeneous systems. But proposed methods are relatively complex and not easily adapted for control plane requirements. To cope with this problem, we account for the non-linear effects (by the means of φnl) due to the SMF and LEAF separately and then combining thembefore computing the Q-factor balancing non-linear effects accumulated in LEAF links [8]. The interest of this approach is its simplicity: compared to the previous accumulation of physical impairments to estimate the Q-factor, we have just to distinguish the φnl due to the two fiber types and then translate LEAF non-linear effects with respect to SMF ones.

The simulation and experimental set-ups to emulate different scenarios of SMF and LEAF fiber successions use the same scheme, represented in Fig. 6. NRZ 10 Gb/s transmission is studied. Two combs of 100 GHz-spaced channels are

encoded with 223-1 long pseudo-random binary sequences and are interleaved. The resulting WDM comb is passed into an acousto-optic switch and injected into a dual-arm recirculating loop [15] The loop has two possible directions (one per each fiber type) and enables the emulation of a mix of fiber-types in a path; the choice of loop direction depends on the state of the two other acousto-optic switches (Switches 2 and 3 in Fig. 6). Wavelength-selective switches (WSS) are placed after each link to equalize channel power levels. The SMF loop is made up of three 100 km length spans, whereas the LEAF is made up of three 75 km- and one 100 km-length spans. DCF are placed in double-stage amplifiers and used for pre- and post- compensation to ensure a return to zero after each loop, as explained in Section I.2. At loop output a tunable dispersion compensation module (TDCM) is used for performance optimization.

Figures 7(a) and 7(b) show the different results of OSNR penalties obtained with measurement and simulation. As in the previous case, we compared the measured and calculated OSNR penalties for the same link set-up. We observed, for this case, that the QoT estimation obtained by the use of simulation presents an off-set compared to the estimation obtained by experimental measures. This is due to the difficulty to emulate by simulation the highly non-linear propagation in LEAF fibers. Hence, we have to introduce in this case an offset for calibrating the QoT estimator obtained by simulation and for using it to estimate the performance of deployed networks.

0 0.5 1 1.5Non linear phase ( Rad )

LEAF 3 dBm per channel

SMF 3 dBm per channel

LEAF SMF

0 0.5 1 1.5Non linear phase ( Rad )

LEAF SMF

0 0.5 1 1.5Non linear phase ( Rad )Non linear phase ( Rad )

LEAF SMF

LEAF 3 dBm per channelLEAF 3 dBm per channel

SMF 3 dBm per channelSMF 3 dBm per channel

LEAF SMFLEAF SMF

LSLS SLSL

LLSSLLSS

0 0.5 1 1.5

-Total non linear phase (rad) (b)

LEAFSMF

Other simbols, mix cases

0 0.5 1 1.5

-Total non linear phase (rad) (b)

LEAFSMF

Other simbols, mix cases

Figure 7: Example of experimental (a) and simulation (b) OSNR penalties as a function of non-linear phase for a mix of fiber transmission.

Bit Error Rate(BER)

Variable Noise Generator

OpticalAttenuator

Switch 1

Input WDM (50GHz spaced)10.7Gbits/s NRZ

100 km SMF

Switch 3

OpticalFilter

75 km LEAF

PolarizationScramblerWSS 2

OpticalAttenuator Switch 2

PolarizationScrambler

CD CompFiber

CD Comp Fiber CD Pre CompFiber

CD Pre CompFiber

100 km SMF

100 km LEAF

Bit Error Rate(BER)

Variable Noise Generator

OpticalAttenuator

Switch 1Switch 1

Input WDM (50GHz spaced)10.7Gbits/s NRZ

100 km SMF

Switch 3

OpticalFilter

75 km LEAF

PolarizationScramblerWSS 2

OpticalAttenuator Switch 2

PolarizationScrambler

CD CompFiber

CD Comp Fiber CD Pre CompFiber

CD Pre CompFiber

100 km SMF

100 km LEAF

Figure 6: Scheme of the experimental set-up of the double loop used to emulate different scenario of mix SMF and LEAF fiber successions.

IV. IMPACT OF ROUTING ON TRANSMISSION PERFORMANCE

As mentioned before, the proposed QoT estimators (such as [2] and [3]) usually predict the performance considering a worst case propagation: in the WDM optical comb all channels are lit and they propagate together along the path (all channels have the same input and output nodes). But in a practical network this situation is not accurate, because intermediate optical switch nodes enable the connection between different directions and the insertion/extraction of channels. This means that optical switches (OXC) route adjacent channels towards different output ports, varying the distance traveled by the signals and XPM penalties on a channel under examination. Standard point-to-point experimental set-ups cannot emulate changing neighboring channels because they cannot extract or insert channels after propagating a variable number of links. To allow for such emulation, two experimental set-ups are proposed, [16] and [17].

A. Study of propagation distance distribution

The study presented in [16] uses a recirculating loop with a non-uniform transmission distance distribution that allows for multiple groups of channels within the loop to propagate for different distances (number of circulations). To achieve this functionality, 8 separately triggered electro-optic single-pole double-throw loop switches are used to alternate between loading and propagating signals. Each channel can be assigned to any of the 8 loop switches using the 1x9 input wavelength selective switch (one port is not used). A 1x8 splitter at the loop output sends all channels to the 8 loop switches and each channel is then restricted to the appropriate loop switch during circulation by the routing assignment in two 1x4 WSS units combined with a 3 dB coupler to create 1x8 switch functionality. The loop output coupler is placed at the end of the loop and the switch triggers are timed to coincide with the loop propagation round trip time. In this way, the performance after the nth round trip is measured during the loading time andthe signals only propagate for at most n round trips. Depending on their relative propagation times, the channel groups will load at different times. A separate trigger is used to gate the BERT and OSA within the propagation time of a given set of channels. Using this triggering, an average is taken over the impact of the other groups loading at different times. Using this triggering which is asynchronous between certain channel groups samples the different add/drop combinations.

The transmission fiber consists of 4 spans of SSMF fiber with nominal lengths of 85, 83, 93, and 68 km. The dispersion map uses pre-compensation of -900 ps/nm and a residual dispersion per span of +100 ps/nm. The dispersion at the add-drop nodes (after 1 round trip) is returned to zero. Forty 100 GHz spaced C band sources are muxed together and modulated at 10.7 Gb/s using NRZ on-off keying as above.

Figure 8: Scheme of the experimental set-up used to emulate the diverse channel propagation distances.

1530 1535 1540 1545 1550 1555 1560 156510

24 8 hops 7 hops 6 hops 5 hops 4 hops 3 hops 2 hops 1 hop Asgnmt Meas.

Wavelength (nm)Figure 9: QoT performance estimate and measurement (green stars) as a function of number of loop round trips (hops) 1-8.

A comparison of the Q-factor measured after propagateing channels along a different number of links and estimated by the use of the QoT-estimator developed in section 2 is depicted in Fig. 8. We observe an agreement of the Q-factor behavior, as it is represented in Fig. 9. Some uncertainty is introduced at very long distances due to the large nonlinear phase.

B. Study of the XPM due to time varying neighbors

In the following we investigate the impact of changing neighbors using simulations and experimental measurements, respectively, to estimate the decrease of XPM penalties in the required OSNR.

1) Simulation set-up In [19] the simulation set-up is similar to the one described

in section II.1, where 19 10.7 Gb/s NRZ channels propagate over 4 span links. The evolution of the central channel (1550.12 nm) is investigated after each OXC. Unlike the previous approach, the potential change of adjacent channels is included. As signal degradation due to non-linear effects depends on the accumulated chromatic dispersion per span or link between adjacent nodes [7], two dispersion maps are studied: one with full chromatic dispersion compensation between adjacent nodes (detrimental for XPM effects between adjacent channels: since the relative position of information sequences between adjacent channels is identical after each node due to zero accumulated dispersion, i.e. identical group

delays, then the impact of XPM may grow resonantly after each node) and another allowing a residual dispersion between adjacent nodes of +90 ps/nm. As expected, when the dispersion map has regular full inline dispersion compensation, the OSNR penalties are higher (Fig. 10(a)) compared to the OSNR penalties observed when the dispersion map has a positive return at the end of each link (Fig. 10(b)).

We also observe that higher penalties occur when neighboring channels propagate together with the observed channel along the path, which can be attributed to XPM. Conversely, partial renewal of channels at nodes may break XPM resonant accumulation. The 0% channel change probability represents the case of a point-to-point transmission and show the highest penalties for the dispersion map. After dimensioning a network as in [19], the average number of neighbor changes that a channel notices along its propagation is 50%, while only 0.2% of channels do not notice any change

in their neighbors [19]. This observation further illustrates the need for alternatives to point-to-point transmission experiments for the study of transparent mesh networks.

2) Experimental set-up In [20] the same experimental set-up proposed in Fig. 11

has been proposed to emulate the neighboring channel variations in a transparent mesh network. The experimental set-up is made of a circulating loop of three 100 km-length SMF spans and a coupler emulates the transit in an optical switch and the resulting changes in neighboring channel configurations. As in the previous simulation work, the residual dispersion per span is set to be 0 or +90 ps/nm (called hereafter RCD0 and RCD90, respectively), while the per span residual dispersion is +100 ps/nm.

2 4 6 8 10

Number of cascaded OXCs

baseline

RCD90RCD90 Channel Change Probability

40 %60 %

0 %20 %

2 4 6 8 10

baselinebaseline

40 %60 %

0 %20 %

Channel Change Probability

40 %60 %

0 %20 %

2 4 6 8 10

baselinebaseline

40 %60 %

0 %20 %

40 %60 %

0 %20 %

2 4 6 8 10

baselinebaseline

40 %60 %

0 %20 %

40 %60 %

0 %20 %

2 4 6 8 10

baseline

40 %60 %

0 %20 %

2 4 6 8 10

baselinebaseline

40 %60 %

0 %20 %

40 %60 %

0 %20 %

2 4 6 8 10

baselinebaseline

40 %60 %

0 %20 %

40 %60 %

0 %20 %

2 4 6 8 10

baselinebaseline

40 %60 %

0 %20 %

40 %60 %

0 %20 %

(b) Figure 10: OSNR penalties for channels propagating in a transparent way as a function of the channel change probability. (a) The impact of XPM is shown when the dispersion map has a return to zero at each node (RDC0); (b) while no XPM is present if the dispersion map present a positive residual dispersion (RCD90).

Results obtained in these studies are similar to results observed by simulation. From Fig. 12(a) a decrease of XPM penalties is observed if the probability of a neighboring channel switch increases: after 7 loops penalties exceed 4 dB when all neighbors propagate together, but fall to 1 dB when neighbors are switched at each intermediary WSS. We also observe that as in deployed backbone networks the probability of a neighboring channel being changed is around 50%, the QoT estimation is underestimated of about 1 dB. From Fig. 12(b) we observe that with a dispersion map having a Dres per node of +90 ps/nm (no re-synchronization of adjacent channels) the XPM effects are

insignificant: penalties are not due to the number of neighboring channel switches (0% and 100% probability nearly have the same behavior) because the bits of the adjacent channels do not re-synchronize anymore. Moreover, we tested two different modules to post compensate (-500 ps/nm) or not (0 ps/nm) the residual CD that is no longer nulled after the propagation (RCD90). We observe that in this case, there is no additional margin on the estimated Q-factor and performance depends on the post CD-compensation at the receiver: better performance is observed if tunable compensation is done.

I. CONCLUSIONS

In this paper we presented diverse experimental set-ups to emulate optical network heterogeneity. Mesh opticalnetworks are more difficult to emulate than traditional point-to-point systems because of the variety of propagation scenarios. The diversity of paths among channels propagating together in a fiber link influences theaccumulation of physical impairments. The heterogeneity can also be related to the characteristics of devices installed in the links and nodes. To emulate such heterogeneous mesh networks, a double loop was realized and used to validate a quality of transmission (QoT) estimator originally generated through simulations. Simulations were used to create

switch 1

Bit Error Rate(BER)

WDM comb21 channels

10.7Gb/s, NRZ(50GHz spaced)

variable Noise generator

OpticalAttenuator

DCFWSS 1 CD comp.

Polarizationscrambler

Opticalfilter

100 km SMF

CD postcomp. fibre

switch 2

switch 3

switch 1

Bit Error Rate(BER)

WDM comb21 channels

10.7Gb/s, NRZ(50GHz spaced)

variable Noise generator

OpticalAttenuator

DCFWSS 1 CD comp.

Polarizationscrambler

Opticalfilter

100 km SMF

CD postcomp. fibre

switch 2

switch 3

Figure 11: Experimental set-up to emulate impact of neighbor changes in a transparent mesh network.

2 4 6 8 10

baseline

RCD0RCD0

(dB/0.2

for BER

Number of cascaded OXCs2 4 6 8 10

baselinebaseline

RCD0RCD0

(dB/0.2

for BER

2 4 6 8 10

baselinebaseline

RCD0RCD0

(dB/0.2

for BER

2 4 6 8 10

baselinebaseline

RCD0RCD0

(dB/0.2

for BER

baseline

RCD0RCD0

(dB/0.2

for BER

baselinebaseline

RCD0RCD0

(dB/0.2

for BER

2 4 6 8 10

baselinebaseline

RCD0RCD0

(dB/0.2

for BER

2 4 6 8 10

baselinebaseline

RCD0RCD0

(dB/0.2

for BER

different network states (various combinations of physical effects impacting the signal transmission). But, such an estimator needs to be validated because models might neglect some effects present in transmission.

In this paper we presented experiments that enabled the study of networks having uncertainties in dispersion maps,

heterogeneity of fiber types (mix of SMF and LEAF) and also varying neighbors propagating with the observed channel. For all these scenarios we presented simulation and measured channel performance and we have shown that in some cases the experimental validation was necessary to improve the accuracy of the proposed estimator.

(a)(a)0

0 2 4 6 8

0% 25%50% 100%

RCD0RCD0RCD0

(a)(a)0

0 2 4 6 8

0% 25%50% 100%

RCD0RCD0RCD0

RCD90RCD90RCD90

0 2 4 6 8Number of cascaded OXCs

Channel Change ProbabilityChannel Change Probability0%: post CD0 post CD-500

100%: post CD0 post CD-500

(b)(b)

RCD90RCD90RCD90

(b)(b)

Figure 12: Measure of OSNR penalties for channels propagating in a transparent way as a function of the channel change probability. (a) The impact of XPM is shown when the dispersion map has a return to zero at each node (RDC0); (b) while no XPM is present if the dispersion map present a positive residual dispersion (RCD90).

REFERENCES

[1] B. Lavigne, F. Leplingard, L. Lorcy, E. Balmefrezol, J.-C. Antona, T. Zami, D. Bayart, “Method for the Determination of a Quality-of-Transmission Estimator along the Lightpaths of Partially Transparent Networks”, in Proceedings ECOC 2007, September 2007.

[2] F. Leplingard, T. Zami, A. Morea, N. Brogard, D. Bayart, “Determination of the Impact of a Quality of Transmission Estimator Margin on the Dimensioning of an Optical Network”, in Proceedings IEEE/OSA OFC 2008, March 2008.

[3] S. Pachnicke, J. Reichert, S. Spalter, E. Voges, “Fast analytical assessment of the signal quality in transparent optical networks”, Journal of Lightwave Technology, Vol. 24, No. 2, February 2006, pp. 815-824.

[4] A. Morea, N. Brogard, F. Leplingard, J.-C. Antona, T. Zami, B. Lavigne, D. Bayart , “QoT function and A* routing: an optimized combination for connection search in translucent networks”, Journal of Optical Networking, Vol. 7, No. 1, January 2008, pp. 42-61.

[5] G. P. Agrawal, “Non-linear fiber optic” (Academic, 2001). [6] ITU-T Recommendation G.652, “Recommendation G.652:

Characteristics of single-mode optical fibre and cable”. [7] E. Desurvire, D. Bayart, B. Dethieux and S. Bigo, “Erbium-Doped

Fiber Amplifiers, Device and System Developments,” Wiley & Sons Interscience, New York, 2002 (Chapter 7).

[8] J.-C. Antona, S. Bigo, Physical design and performance estimation of heterogeneous optical transmission systems, C.R. Physique 9 (2008) 963-984

[9] ITU-T Recommendation G.691, “Optical interfaces for single-channel STM-64, STM-256 and other SDH systems with optical interfaces,” ITU, 200).

[10] J. D. Downie and A. B. Ruffin, “Analysis of signal distortion and crosstalk penalties induced by optical filters in optical networks,” Journal of Lightwave Technologies, Vol. 21, No. 9, pp. 1876-1886, September 2003.

[11] J. C. Antona, S. Bigo, J. P. Faure, “Nonlinear cumulated phase as a criterion to assess performance of terrestrial WDM systems,” in Proceedings of IEEE/OSA OFC 2002, pp. 365-366.

[12] ITU-T Recommandation G. 655, “Characteristics of a non-zero dispersion shifted single-mode optical fibre and cable”.

[13] D. Breuer, N. Hanik, C. Caspar, F. Raub, G. Bramann, M. Rohde, E.-J. Bachus, S. McLeod, M. Edwards, “Mixed Fiber Infrastructures

in Long Haul WDM-Transmission”, Journal of Optical Communications, Vol. 25, No. 1, pp 10-13.

[14] S. Pachnicke, N. Hecker-Denschlag, S. Spalter, J. Reichert, E. Voges, “Experimental verification of fast analytical models for XPM-impaired mixed-fiber transparent optical networks,” Photonics Technology Letters, Vol.16, No. 5, May 2004, pp 1400-1402.

[15] P. Peloso, M. Prunaire, L. Noirie, and D. Penninckx “Optical transparency of a heterogeneous pan-European network,” J. Lightwave Technology, vol. 22, no. 1, pp. 242–248, Jan. 2004

[16] D. C. Kilper, D. Bayart, S. Chandrasekhar, A. Morea, S. K. Korotky, F. Leplingard, “Mesh Network Transport Experiments Using a Distributed-Distance Circulating Loop,” in Proceedings IEEE ECOC 2008, paper We.3.D.5.

[17] T. Zami, P. Henri, L. Lorcy, C. Simonneau, “Impact of the optical routing on the transmission in transparent networks,” in Proceedings IEEE ECOC 2009, paper 1.5.2.

[18] J.-C. Antona, M. Lefrançois, S. Bigo, G. Le Meur, “Investigation of Advanced Dispersion Management Techniques for Ultra-Long Haul Transmissions”, ECOC’05, Mo.3.2.6, Sept 2005.

[19] T. Zami, A. Morea, N. Brogard, “Impact of routing on the transmission performance in partially transparent optical network,” in Proceedings of IEEE/OSA OFC 2008, paper JThA50.

[20] T. Zami, P. Henri, L. Lorcy, C. Simonneau, “Impact of the optical routing on the transmission in transparent networks”, in Proceedings IEEE ECOC 2009, paper 1.5.2.

[21] S. Chandrasekhar, X. Liu, “Impact of Channel Plan and Dispersion Map on Hybrid DWDM Transmission of 42.7-Gb/s DQPSK and 10.7-Gb/s OOK on 50-GHz Grid”, IEEE Photonics Technology Letters, Vol. 19, No. 22, November 2007, pp. 1801-1803.

[22] A. Morea, D. C. Kilper, I. S. Lin, F. Leplingard, S. Chandrasekhar, T. Zami, J.-C. Antona, “testbed methods to study physical layer path establishment in long haul optical wavelength switched networks,” in Proceedings IEEE ICTON 2009, paper Tu.C2.4.

All-optical Label Swapping Techniques for Optical Packets at Bit-rate Beyond 160 Gb/s

Nicola Calabretta, Hyun-Do Jung, and Harm Dorren COBRA Research Institute, Eindhoven University of Technology, Eindhoven, The Netherlands.

E-mail: n.calabretta@tue.nl , Email: {h.d.jung, h.j.s.dorren}@tue.nl

Abstract— In this paper two different paradigms to realize a scalable all-optical packet switch with label swapping will be presented. All the functions required for switching the packets are based on all-optical signal processing without any electronic control. This allows very low latency and potential photonic integration of the systems. We report for both techniques experimental results showing the routing operation of the 160 Gb/s packets and beyond. We will discuss and compare both techniques in term of devices and bit-rate scalability, latency, power consumption, power penalty performance and cascadability as key parameters for the realization of an all-optical packet switch.

Index Terms— Optical packet switching, optical signal processing, label processor, label rewriter, label swapping, semiconductor optical amplifier.

I. INTRODUCTION

All-optical packet switching has been proposed as a technology to solve the bottleneck between the fibre bandwidth and the electronic router capacity by exploiting high speed and parallel operation of all-optical signal processing. Moreover, photonic integration of the optical packet switch potentially allows for a reduction of volume, power consumption and costs. In all-optical packet switches the optical packets are routed based on the address information that is encoded by the attached labels. The optical packet is stored (delayed) in the optical domain for the time required to the label processor to process the address and provide a routing signal for routing all-optically the stored packet.

To exploit the benefit of photonic technology to miniaturize and decrease the power consumptions of the system, photonic integration of the all-optical packet switch depends on the capability to integrate the label processor and the optical delay related to the latency of the label processing. This imposes stringent constraints on the latency time of the label processor. Indeed, integrated delay lines using an InP photonic waveguides have around 2 dB/cm of optical losses. One centimeter of waveguide provides a delay of 100 ps. If the latency of the label processor is in the order of 1 nanosecond, integration of such delay exhibits a total waveguide loss of 20 dB, which is unpractical. Therefore, high speed operation of the label processor (< 100 ps) is a must to allow photonic integration of the packet switch system.

Moreover, scalability of the label processor with the number of labels (or the number of label bits) is crucial too.

Several solutions were presented to implement an all-optical packet switch node. In [1-5], the addresses were processed in the electrical domain while the payload is stored in the optical domain. The electrical label processing drives the optical switches for routing the optical packets. However, electronic label processing and new label rewriting requires no trivial optoelectronic per-packet based clock recovery, and introduces long processing latency in the order of tens of nanoseconds which prevents the integration of the system. All-optical packet switch employing all-optical label processor were investigated in [6-12]. Mainly these works employed optical correlators, which recognize the labels, and set/reset optical flip-flops to store the information for the duration of the packet. However, as the number of addresses, of the Wavelength Division Multiplexing (WDM) channels carried by each fiber, and of the packet data rate increase, photonic integration, high speed operation, low latency, and scalability of the label processor remain key-issues to be solved. Solutions employing 2N optical correlators and 2N optical flip-flop to process the addresses may prevent photonic integration.

Our research focuses on the realization of an all-optical packet switching system that is scalable and suitable for photonic integration. We present two all-optical packet switching techniques [13, 14] that utilize all-optical signal processing to implement the label processor and the label rewriter. The two all-optical label swapping (AOLS) techniques are based on two different paradigms. One is based on wavelength routing switching [13] and the other one on space routing switching [14]. Both techniques employ scalable and asynchronous label processor and label rewriter capable to process optical in-band labeling addresses. We demonstrate a 1x4 all-optical packet switch based on both techniques. For both techniques we report experimental results showing the routing operation of the 160 Gb/s packets based on the processed in-band address information, and all-optical label erasing and new label insertion operation. Based on the experimental results, we discuss and compare both techniques in term of devices and bit-rate scalability, latency, power consumption, power penalty performance and cascadability as key parameters for the realization of an all-optical packet switching node.

The paper is organized as follows. In Section II, we present the all-optical packet switch architecture, introducing the main all-optical functions required to accomplish the AOLS. In Section III and section IV, we present two all-optical packet switching techniques that utilize all-optical signal processing to implement the label processor and the label rewriter. The two techniques are based on wavelength routing switching and on space routing switching. Section V provides a comparison between the two techniques. Finally, we summarize and discuss the main results in the conclusions section.

II. AOLS ARCHITECTURE

Figure 1 illustrates the all-optical packet switch based on label swapping technique. The input packet format is also reported in figure 1. The input packets consist of a 160 Gb/s payload, with a pulse duration of 1.6ps making the 20 dB bandwidth of the payload to be 5nm. The packet address information is encoded by in-band labels. With this we mean that the wavelengths of the labels are chosen within the bandwidth of the payload. We encode addresses by combining different labels. Each label is On-Of-Keying (OOK) encoded and has a binary value: the label value is ‘1’ if the label is attached to the payload, the label value is ‘0’ if no label is attached to the payload. Thus, by using N in-band label wavelengths, 2N possible addresses can be encoded, which makes this labelling technique highly scalable within a limited bandwidth.

Figure 1. Packet format in the time and spectral domain, and all-optical packet switch configuration.

We have used simulation based on Matlab to calculate the number of possible labels that can be allocated in the spectrum of the payload bandwidth [12]. Note that by using filters with bandwidth narrower than 0.1 nm more than 10 labels can be allocated in the payload bandwidth, which means 210 encoded addresses. Moreover, if the payload data rate increases above 160 Gb/s (i.e. 320 or

640 Gb/s), a larger number of labels can be allocated in the payload spectrum. Thus, the proposed labeling technique scales well with the packet data rate. Other advantages of the in-band labeling are that the labels can be extracted by passive wavelength filtering. Moreover, by using a label that has the same time-duration as the payload makes the use of optical flip-flops redundant, and allows to handle packets with variable lengths in an asynchronous fashion. In the experiment, we encode 4 addresses by using two in-band labels. Figure 1 shows packets carrying different addresses and the corresponding representation in the spectral domain. The all-optical packet switch is based on label swapping technique. In the label swapping technique, the input labels have only a local meaning. The input labels are used to provide the packet’s routing information. New labels should be generated and attached to the packet payload before that the packet outputs the switch. To perform the label swapping and routing of the packet, we utilize four all-optical functions as shown in figure 1: label extraction/erasing, label processing, label rewriting, and switching and labels insertion. The packet address encoded by the in-band labels is extracted/separated from the data payload by the label extractor/eraser. The data payload is optically delayed for the time required to the label process to provide a routing signal, before being fed into the switching and labels insertion. The labels are all-optical processed by the label processor and label rewriter. The label processor provides a routing signal according to the input labels. The routing signal at unique wavelength has a time duration equal to the packet time. The wavelength of the routing signal is used to drive the switching and labels insertion. Simultaneously, the label rewriter provides the new labels, which have a time duration equal to the packet duration. Moreover, the wavelengths of the new labels are selected so that they are in-band with the bandwidth of the converted payload. The new labels are attached to the switched. It is worth to note that since the label processor and label rewriter operate ‘on the fly’, the time delay required to store the payload is very short. This may allow photonic integration of the whole packet switch system. Moreover, as the routing signal and the new labels produced by the label processor and label rewriter have a time duration equal to the packet time, the presented system can handle packets with variable length. An example of self-routing table for two labels addresses is reported in figure 2. For each input labels combination, a routing signal at distinctwavelength and a new combination of labels should beprovided by the label processor and the label rewriter,

Figure 2. Routing table used in the label swapping.

respectively. Figure 2 reports also the corresponding optical spectra of the routing signal and new labels for different input labels combination. Note that the wavelengths of the new labels should be within the 5 nm band of the payload.

III. AOLS BASED ON WAVELENGTH ROUTING SWITCH

The first AOLS technique is based on wavelength routing switching [13]. To perform the label swapping and routing of the packet, we utilize four all-optical functions as shown in figure 3. The input packet is firstly processed by the label extractor/eraser, which consists of fiber

Figure 3. AOLS based on wavelength routing switch.

Bragg gratings (FBG) centered at the labels wavelengths. While the labels are reflected by the FBGs, the packet payload can pass through the label extractor/eraser before to enter the wavelength converter. The continuous wave (CW) routing signal that is needed for wavelength conversion is provided by the label processor. The optical power of the extracted labels is used to drive the label processor. The label processor receives also as input 2N

CW bias signals at different wavelengths λ1…λ2N. The

wavelengths of the CW-signals are chosen according to the self-routing table and represent the wavelengths at which the payload will be converted. The label processor consists of a cascaded of N pairs of periodic filter and optical switch. The periodic filter has one input and two outputs. The optical switch has two inputs and one output. The two outputs of the periodic filter have complementary wavelength transfer functions figure 3. Moreover, each of the N periodic filters has different period as also shown in figure 4. In particular the bandwidth (BW) of the i-th filter is equal to BWi=2(i-1) x BWch, with i=1,…, N and BWch, the bandwidth of the single CW-signal. Each of the 1 x 2 periodic filter separates (in wavelength) half of the input CW-signal to output port 1 and the other half of the input CW-signals at the output port 2. The 2 x 1 optical switch selects the CW-signals of port 1 or port 2 based on the value of the label information. Therefore, the output of each pair of periodic filter and optical switch consists of half the number of CW-signals. Thus, after the first stage, the 2N

CW-signals becomes 2N /2 = 2N-1. Therefore, after cascading N pairs in which each optical switch is driven by the corresponding label, a distinct CW-signal is selected. This CW-signal at distinct wavelength has a time duration equal to the packet and represents the

Figure 4. Label processor set-up.

routing signal to which the payload will be converted. Note that the processing is performed entirely in the optical domain. By implementing the optical switches by means of very fast Semiconductor Optical Amplifier – Mach-Zhender Interferometer (SOA-MZI) devices, label processing with only tens of picoseconds of processing time can be possible. Moreover, as no synchronization is required in the scheme, and the routing signal at the output of the label processor has the same duration as the packet payload, the system can handle packets with variable lengths. For each input labels combination, the label processor provides a routing signal according to the input labels. The routing signal at unique wavelength has a time duration equal to the packet time. The wavelength of the routing signal represents the central wavelength at which the 160 Gb/s data payload will be converted by means of wavelength conversion [15, 16]. Simultaneously, the label rewriter, which is based on the same operation principle of the label processor, provides the new labels, which have a time duration equal to the packet duration. Moreover, the wavelengths of the new labels are selected so that they are in-band with the bandwidth of the converted payload. The new labels are attached to the wavelength converted payload. The packet with the new labels is routed by means of an Array Waveguide Grating (AWG) to distinct output ports of the packet switch, according to the central wavelength of the converted payload as shown in Figure 5.

Figure 5. Switched packets at the four output ports.

We set the CW-signals according to the label swapping table reported in figure 2. Figure 6b shows the spectrum of the payload signal after label extraction. As compared with figure 6a, the label was erased. Based on two-labels

Figure 6. Optical spectra of the packet recorded at a) before the label extractor; b) after the label extractor; c) wavelength converted payload with attached the new label; d) after the label extractor/eraser of the receiver node.

combination and according to the self-routing table, the label rewriter gave four sets of new labels. Figure 7shows the BER performance at different position of the two nodes system. The BER measurements were performed in a static operation by using a 160 Gb/s PRBS 231-1 data payload and fixing one address (old label ‘0 1’). The label extractor causes a penalty of less than 0.5 dB compared to the back-to- back payload.After the wavelength conversion, error-free operation was obtained with 5.5 dB of penalty. As reference we also reported the 160 Gb/s back-to-back wavelength converted, which has 4 dB of power penalty. The additional 1.5 dB penalty compared with 160 Gb/s back-to-back wavelength conversion can be ascribed to the pulse broadening by the label extractor which affects

Figure 7. BER measurements and eye diagrams at different points of the system. Time scale is 2 ps/div.

the wavelength conversion performance. The switched packet with the new label (1, 0) (see spectrum in figure 6c) was then fed into the receiver node. The optical spectrum of the packet of the new labels.after the label extractor of receiver node is reported in figure 6d. The power penalty after the label extractor is 0.5 dB. This results in a limited power penalty caused by the extraction/ insertion.

IV. AOLS BASED ON SPACE ROUTING SWITCH

The schematic of the AOPS is shown in figure 8. The AOPS consists of a label extractor/eraser, an optically controlled tunable laser (OCTL), and optical gates for payload switching and label rewriting. The input packets are firstly processed by the label extractor/eraser, which consists of two fiber Bragg gratings (FBG) centred at λL1

and λL2, respectively. The data payload passes through the label extractor/eraser and is broadcasted into the optical gates. The two labels are reflected by the FBGs and fed into the label processor via optical circulators. The labels optically control the output wavelength of the OCTL. The OCTL output acts as a control signal for one of the SOA-MZI based optical gates. These optical gates have two functions. Firstly, they route the packet payload according to the routing table. Secondly, they rewrite the new labels. The OCTL consists of four cw-lasers, two SOA-MZIs and 2 AWGs [14]. The cw-signals are pair-wise fed into the two inputs of SOA MZI 1. The control signal of SOA-MZI 1 is label 1. Thus the presence of label 1 selects two of the cw-signals. Conversely, if label 1 is not present, the other two cw-signals are selected. The two cw-signals that output SOA-MZI 1 are separated by an AWG. Each of the separated cw-signals is fed into one of the two inputs of SOA-MZI 2. The control signal of SOA-MZI 2 is label 2. Thus the presence of label 2 selects one of the two cw-signals that act as a control signal for the optical gates. Each of the four cw-signals can be selected by a combination of the two labels. Both the payload and the new cw-label are fed simultaneously in the SOA–MZI gate that is controlled by the OCTL output. If a control signal is present, the SOA-MZI gates both the packet payload together with the new label to the output. Conversely, the gate-output is blocked. The operation of the gate guarantees that the payload and the new label have the same duration at the gate output. Figure 9 shows the switched packets at the four outputs of the optical packet switch. It is worth to note that since the label processor and label rewriter operate ‘on the fly’, the time delay required to store the payload is very short. This may allow photonic integration of the whole packet switch system. Moreover, as the routing signal and the new labels produced by the label processor and label rewriter have a time duration equal to the packet time, the presented system can handle packets with variable length. To evaluate the performance, the switched packets are fed into a receiving node, consisting of a label extractor (only the payload is evaluated), a 160-to-10 Gb/s demux, and a 10 Gb/s detector. Fig. 10 shows the BER curves. As reference we report the BER curve of the back-to-back (b-t-b) 160 Gb/s payload. The BER curve of the switched packet at Output 2 (no new label inserted) shows error-free operation with 1 dB of power penalty. We also report the BER curve of the switched packet at Output 3, in which a new label (‘01’) is inserted. An additional power penalty < 0.5 dB was measured compared to the case without label insertion. This indicates that the switch with label rewriting introduces very small penalties. As a final result we report in Fig. 10 the eye diagrams of the b-to-b payload

and the switched payload at 320 Gb/s. Although the eye diagram gives only qualitative information, the clear open eye suggests that error-free operation at 320 Gb/s is feasible.

Fig. 9 Measured traces showing 160 Gb/s payload and the output traces of the AOPS. The vertical scale is in mV.

Fig. 10 BER curves. Time scale eye diagrams 1ps/div.

V. COMPARISON BETWEEN THE TWO AOLS TECHNIQUES

Device scalability: The AOPS should be scalable in terms of number of input and output ports and number of components. It is important that these large AOPS can still be controlled with a limited amount of signals and that the number of control signals scales efficiently with the number of input and output ports. Finally, it is important that the switch introduces low signal degradation. The AOPS based on wavelength routing switch scales better than AOPS based on space switch in terms of number of components. This is due to the fact that the label processor (and label rewriter) and the wavelength converter requires 1+(log2 N) active components (see scheme in section 3), while in the space switching N active components (see scheme in section 4) are required. The main limitation of the label rewriter for the AOPS based on wavelength routing switch is the Optical Signal to Noise Ratio (OSNR) degradation with the increase of the number of labels. On the contrary, in the space switch the OSNR degradation is much reduced since the label rewriter and the switching are implemented by a single active component. Therefore, for AOPS with a limited number of input/output port (limited number of labels), the wavelength routing technique is preferable due to the limited number of active components. For large input/output ports, the space switch is preferable. Bit-rate scalability: The AOPS should be able to operate at data rate beyond 160 Gb/s. For bit-rate beyond 160 Gb/s, the AOPS space switch outperforms the AOPS wavelength routing switch mainly because the capability to operate the wavelength converter with data rate beyond 160 Gb/s with acceptable power penalty. Error-free operation was attained in both tecniques. However, in the space switching technology (section 4) the penalty was 1.5 dB compared to the BER measured in back-to-back configuration. It is expected that at higher bit rate, this penalty will increase but should be acceptable for cascadability of the AOPS. On the other hand, the technique based on wavelength routing switch already performs more than 5 dB of penalty and this will even be higher at data rate beyond 160 Gb/s, at least with this wavelength conversion technique.

Figure 8. AOLS based on space routing switch.

0 0 0 0 0

Latency: The latency is due to the label processor in the wavelength routing switch case, and to the OCTL in the space switch. Both devices introduce the same amount of latency. Node cascadability: Cascadability of the AOPS is mainly limited by the power penalty introduced by the switching technique. In the AOPS based on wavelength routing switch the measured BER penalty was 6.5 dB (section 3), of which 5 dB is due to the wavelength converter operation. In the space switch we have recorded a penalty of 1.5 dB (section 4). The low power penalty allows for cascading the AOPS for several nodes before that the accumulated nonlinearities and degradation of the OSNR. Those considerations lead to prefer the AOPS based on space switching for multi-hops operation. However, more in depth analysis is currently under investigation by using numerical tools. A possible solution for improving the cascadability of the AOPSs is the introduced after the switch (or after a number of hops) of an optical regenerator.

VI. CONCLUSIONS

We have demonstrated a 1x4 packet switch with label swapping based on two techniques. All the required functions to switch the packets and to rewrite the new labels have been implemented in all-optical manner. Both techniques employ a scalable labeling technique that by combining N in-band labels, which wavelengths are within the bandwidth of the payload, can encode up to 2N

possible addresses within a limited bandwidth. The label processing technique requires only N active devises to process ‘on the fly’ the 2N addresses, which makes this technique scalable with the number of addresses. The label processor is based on ‘on the fly’ optical signal processing in SOA MZIs, and on a packet-by-packet basis. This makes extraction of a clock redundant and ensures that the AOPS is suitable for photonic integration and allows very fast operation. This leads to a processing time of few tens of picoseconds, allowing short packet’s guard time. Moreover, being the labels in-band and with a time duration equal to the packet payload, the label processor does not require all-optical flip-flop, operates in asynchronous fashion and can handle packets with variable lengths. We have experimentally measured that label erasing and new label insertion operation introduces only 0.5 dB of power penalty. In terms of latency, note that the latency is due to the label processor in the wavelength routing switch case, and to the OCTL in the space switch. Both devices introduce the same amount of latency. BER measurements on the 160 Gb/s switched packets show error-free operation with a power penalty of 6.5 dB in the case of all-optical packet switch based on wavelength routing switch. For optical switches based on space routing switch, error-free operation with a power penalty of less than 1.5dB was measured. Open eyes indicate that error-free operation for 320 Gb/s payload is possible. Those results indicate that AOPS based on space switching is preferable for multi-hops operation in packet-switched network.

ACKNOWLEDGMENT

This work was supported by the Netherlands Science Foundation (NWO) and Netherlands Technology Foundation (STW) through the NRC Photonics and Vi programs. The authors wish to thank Dr. Eduward Tangdiongga and Dr. Javier Herrera Llorente for help setting the 160 Gbit/s wavelength converter.

REFERENCES

[1] S.J.B. Yoo, ‘Optical packet and burst switching technologies for future photonic internet,’ Journal of Lightwave Technology, vol. 24, pp. 4468-4492, 2006.

[2] D.J. Blumenthal, ‘Optical packet switching,’ Proc. IEEE Lasers and Electro-Optics Society, LEOS 2004, pp. 910-912, 2004.

[3] J.J. Vegas Olmos et al, ‘Simultaneous optical label erasure and insertion in a single wavelength conversion stage of combined FSK/IM modulated signals,’ IEEE Photonics Techn. Letters, vol. 16, pp. 2144-2146, 2004.

[4] D. Klonidis et al, ‘OPSnet: design and demonstration of an asynchronous high-speed optical packet switch,’ Journal of Lightwave Technology, vol. 23, pp. 2914-2925, 2005.

[5] B.R. Koch et al, ‘Payload-envelope detection and label-detection integrated photonic circuit for asynchronous variable-length optical packet switching with 40-Gb/s RZ payloads and 10-Gb/s NRZ labels,’ Journal of Lightwave Technology, vol. 24, pp. 3409-3417, 2006.

[6] J.P. Wang et. al., ‘Demonstration of 40-Gb/s Packet Routing Using All-Optical Header Processing,’ Photonics Technology Letters, vol. 18, pp. 2275-2277, 2006.

[7] F. Ramos et al, ‘IST-LASAGNE: towards all-optical label swapping employing optical logic gates and optical flip-flops,’ Journal of Lightwave Technology, vol. 23, pp. 2993-3011, 2005.

[8] M. Takenaka et. al., ‘All-optical packet switching by MMI-BLD optical flip-flop,’ Proc. Optical Fiber Communication 2006, Anaheim, USA, OThS3, 2006.

[9] H.J.S. Dorren et al.,’Optical packet switching and buffering by using all-optical signal processing methods,’ Journal of Lightwave Technology, vol. 21, pp. 2-12, 2003.

[10] P.K.A. Wai et. al., ‘1 x 4 All-optical packet switch at 10 Gb/s,’ Photonics Technology Letters, vol. 17, pp. 1289-1291, 2005.

[11] P.Seddighian et. al., ‘All-optical swapping of spectral amplitude code labels for packet switching,’ Proc. Photonics in Switching 2007, San Francisco, USA, pp. 143-144, 2007.

[12] J. Herrera et al, ‘160-Gb/s all-optical packet switching over a 110-km field installed optical link,’ Journal of Lightwave Technology, vol. 26, pp. 176-182, 2008.

[13] N. Calabretta et. al., ‘1 x 4 all-optical packet switch at 160 Gb/s employing optical processing of scalable in-band address labels,’ Proc. Optical Fiber Communication 2008, San Diego, CA, USA, PDP 33, 2008.

[14] N. Calabretta et. al., ‘160 Gb/s all-optical packet switching with label rewriting,’ Proc. European Conference on Optical Communications 2008, Brussels, Belgium, Th.3.F.3, 2008.

[15] Y. Liu et al.,’Error-free 320 Gb/s all-optical wavelength conversion using a semiconductor optical amplifier,’ Journal of Lightwave Techn., vol. 25, pp. 103-108, 2007.

[16] E. Tangdiongga et al, ‘Monolithically integrated 80-gb/s AWG-based all-optical wavelength converter,’ IEEE Photonic Technology Letters, vol. 18, pp. 1627-1629, 2006.

Nicola Calabretta received the Bachelor’s and the M.S. degrees, both in telecommunications engineering, from Politecnico di Torino, Turin, Italy, in 1995 and 1999, respectively. In 1995, he visited the RAI Research Center (Italian broadcasting television),

Turin, Italy. In 2004 he received the Ph.D. degree from COBRA Research Institute, the Eindhoven University of Technology, Eindhoven, The Netherlands. From 2004 to 2007 he was working as researcher at Scuola Superiore Sant’Anna University, Pisa, Italy. He is currently with COBRA Research Institute, the Eindhoven University of Technology, Eindhoven, The Netherlands. Dr. Calabretta co-authored more than 120 papers published in international journals and conferences and holds 4 patents. He is currently acting as a Referee for several IEEE and IEE and OSA Journals. His fields of interest are all-optical signal processing for optical packet switching, semiconductor optical amplifier, all-optical wavelength conversion and regeneration, and advanced modulation formats for optical packet switching.

Hyun-Do Jung (M’00) received the B.S. degrees in radio sciences and engineering from Kyunghee University, Korea, in 1999, and the Ph.D. degree in electrical and electronic engineering from Yonsei University, Korea, in 2005. Since Sep. 2005, he has been with the department of electrical engineering, Technical University of Eindhoven, the Netherlands

as a senior researcher where he is involved in EU-Project (FP6 MUFINS and FP7 ALPHA) related to optical packet switching and in-building/access networks. His current research interests include optical systems for communications, optical packet switching network, WDM-PON network and microwave photonics technologies.

H.J.S. Dorren received his M.Sc. degree in theoretical physics in 1991 and the Ph.D. degree in 1995, both from Utrecht University, Utrecht, the Netherlands. After postdoctoral positions he joined Eindhoven University of Technology, Eindhoven, the Netherlands in 1996 where he presently serves as a professor and as the scientific director of the COBRA Research Institute. In 2002 he was also a visiting researcher at the National Institute of Industrial Science and Technology (AIST) in Tsukuba in Japan. His research interests include optical packet switching, digital optical signal processing and ultrafast photonics. Prof. Dorren (co-)authored over 250 journal papers and conference proceedings and currently serves as an associate editor for the IEEE Journal of Quantum Electronics.

Figure 1. Systemic view of integration potential using Silica-on-Silicon hybrid integration technology for developing amplitude or phase modulated transmission links: (a) multi-wavelength, multi-format transmitter, (b) multi-format regenerator, (c) (i) Silica-on-Silicon motherboard, (ii) SOA, (iii)

precision-machined silicon submount fabricated by CIP, (iv) DFB laser and (v) InP modulator developed by HHI.

Tb/s Transmission and Routing Systems Using Integrated Micro-Photonic Components

Efstratios Kehayas and Leontios Stampoulidis

Constelex Technology Enablers 17 Ikoniou Street, Acharnai, 13675, Athens, Greece

Tel. +30 609 5723, Fax +30 210 685 8118, email: ek@constelex.eu

Paraskevas Bakopoulos School of Electrical and Computer Engineering, National Technical University of Athens

Zografou, GR 15773 Athens, Greece Tel. +30 210 772 2057, Fax +30 210 772 2077, email: pbakop@mail.ntua.gr

Abstract—Recent advances in the development of photonic switching and transmission systems that exploit high and low index contrast integration materials are reported. Micro-ring resonators, delayed interferometers and all-optical wavelength converters integrated in Si3N4-SiO2 and silica-on-silicon substrates are used for the regeneration and wavelength routing of amplitude and phase modulated optical signals. Micro-photonics is the key for elegantly squeezing terabits into a few mm2 with optimum yield and at a low development cost. Index Terms—high-speed transmission, optical regeneration, photonic routing, all-optical wavelength conversion, photonic integration

I. INTRODUCTION

Today we are witnessing a resurgence of the growth and increasing customer demand for capacity in optical networks. A major factor for this growth is an unprecedented deployment of optical access networks worldwide for providing ample bandwidth to the end-user

in the >50Mb/s region. In this rationale, the growth rates of end-users take new meaning, compared to those in 2000. Before the telecom bubble, penetration rates in excess of 100% reported, fuelled massive technology investments in the core network. However, these growth rates always translated to new users with bandwidth of a few kb/s, since the access network was just not ready. Moreover, broadband applications were still in their infancy and the telecom world was still searching for the “killer application”. The situation is very different today, with each new data connection translating to fast internet with combined voice and video. These new developments in the access networks are now exerting pressure on the metro and core networks that inevitably will always be “out-of-step” with the access networks, given that the genuine and sustainable market drivers for bandwidth are originating from the end-users.

Photonic integrated circuits are expected to play a central role in the development of new hardware to be included in next generation telecommunication systems.

Figure 2. Optical regeneration for OOK and DPSK data (a) system functional blocks, performance evaluation of optical regeneration at 40 Gb/s

with: (b) OOK and (c) DPSK data . Eye diagrams of (i) input data, (ii) degraded data, (iii) data at 2R regenerator output, (iv) BER performance for various levels of degradation (design, fabrication, pigtailing and packaging of devices by CIP Technologies).

In this paper, potential upgrade paths are presented for increasing the throughput of transmission and routing systems utilizing micro-photonics fabricated on material systems that guarantee high integration scale and density. In this rationale, multi-material, multi-functional component integration on a single platform is achieved through optimum combination of monolithic and hybrid integration technologies that play equally important roles. Key milestones on this path are low development costs, integration technology scalability, power consumption and device footprint.

II. MULTI-FORMAT PHOTONIC INTEGRATED TRANSMISSION SYSTEMS

A. The need for integration scale increase Current research efforts for developing new transmission systems in the core network now focus on cost-effective upgrade of transmission capacity using amplitude- and phase- modulated systems: this is where hybrid photonic integration can play a decisive role. The use of “monolithic-on-hybrid” integration approach can enable the increase of integration scale both vertically and horizontally without requiring outstandingly high yields and complex fabrication processes. Combination of high-speed indium phosphide (InP) monolithic arrays (vertical increase) can be on-chip interconnected (horizontal increase) using ultra low-loss silica-on-silicon motherboards [1]. These motherboards or Planar Lightwave Circuit Boards (PLCB) play the role of conventional Printed Circuit Boards (PCB) used in electronics. Using arrayed “all-semiconductor” active components can lead to cost effective, compact and low

power consumption systems. On the other hand, the possibility to integrate these III-V components on a single low-loss circuit board can lead to multi-wavelength, multi-functional devices such as transmitters, receivers and regenerators capable of operating with OOK, DPSK and DQPSK formats, all sharing identical research and development costs. For example an InP-based modulator [2] requires 50% less power to operate and is one order of magnitude smaller than traditional discrete transmitters using LiNbO3. Moreover, considering arrays of such InP modulators hybrid integrated on the PLCB and on-chip interconnected with DFB laser arrays and filtering elements, a terabit capacity photonic integrated circuit becomes a realistic and competitive technology solution.

B. Phase- and amplitude- modulated regeneration systems using silica-on-silicon integration technology

Optical regeneration has the potential for data transparency and more cost-effective mitigation of transmission impairments bypassing the requirement for optical-electrical-optical systems. Hybrid integrated arrays of 2R or even 3R optical regenerators that share development and packaging costs are promising candidates for 40 Gb/s and 100 Gb/s systems, where the use of electronic repeaters becomes challenging, costly and significantly increases power consumption. Exploiting the versatility of the hybrid integration or platform, devices capable of operating with OOK, DPSK DQPSK can be developed using a common fabrication process. In addition, there is also the possibility for multi-functional devices: a single hybrid integrated device can be used for all the data formats Figure 2 shows experimental results of optical 2R regeneration at 40 Gb/s

Figure 3. (a) and (b) TriPleXTM waveguide geometry, (c) SEM image of coupled micro-ring resonators, (d) packaged and pigtailed device. Fabrication by LioniX BV, characterization and pigtailing by XiO

Photonics BV.

Figure 4. Wavelength router implementation using micro-ring resonator arrays

with OOK and DPSK signals using a single device: a hybrid integrated Semiconductor Optical Amplifier Mach-Zehnder Interferometer (SOA-MZI). In order to assess the regenerating capabilities of the device, different levels of amplitude and phase degradations were introduced on the input signal that always resulted in more open eye diagrams and lower BER values for a specific input power to the receiver. Specifically, fig. 2(a) shows the basic functional blocks of the experimental setup, whereas fig. 2(b) and (c) shows experimental results when the SOA-MZI was operated as OOK and DPSK regenerator respectively. Eye diagrams and bit error rate (BER) measurements confirm the regenerative properties of the device, when degraded data are used as test signals.

III. NEXT GENERATION PHOTONIC INTEGRATED ROUTING SYSTEMS

A. Switching systems based on “CMOS photonics”

Increasing the transmission of point-to-point links within the core network cannot translate to useful bandwidth increase in the remaining network chain of metro and access networks, unless core routing systems cope with corresponding upgrade steps. Responding to the call for bandwidth coming from the access networks, system vendors have commercialized new generation of routing systems, following aggressive research and development investments. New terabit routing systems are equipped with 40Gb/s linecards and total throughput of 640 Gb/s. These state-of-the-art routing systems can be upgraded to multi-terabit capacities using multiple interconnected racks in expense of non-linear increase in power consumption, space, size and cooling requirements.

In order for photonics to penetrate into next generation routing systems, several critical milestones need to be achieved: low switching power requirements, small device footprints, low development costs and in the case of routing systems, CMOS-compatible fabrication that may allow merging electronics with photonics on a single platform. The TriPleXTM waveguide technology – developed by Dutch company LioniX BV and now commercialized by XiO Photonics BV – is capable of meeting all these requirements [3]. The TriPleXTM waveguides consist of alternating layers of silicon nitride (Si3N4) and silicon dioxide (SiO2) formed by CMOS compatible low-pressure chemical vapor deposition (LPCVD). Figure 3(a) and (b) show the waveguide geometry; - a low-index SiO2 core is surrounded by a high-index Si3N4 cladding. This material configuration ensures tight waveguide bends and consequently the fabrication of ultra-small structures that can be densely integrated on a single photonic chip. As such, TriPleXTM

can provide photonic chips with a high aggregate throughput that can be efficiently cascaded due to the low insertion (~0.15 dB) and waveguide (<0.006 dB/cm) loss [4]. Figure 3(c) shows a scanning electron microscope (SEM) image of a TriPleXTM ROADM. The component consists of two coupled micro-ring resonators with a ring radius of 50 um and with a capability for independent tuning through integrated heaters. This compact component can be the building block for realizing all the wavelength selective functionalities of a photonic wavelength router offering fine tunability, enhanced chip real estate efficiency and lower cost with respect to mainstream Arrayed Waveguide Grating (AWG) components.

Taking as an example the simplified architecture of Figure 4, a wavelength router consists of:

Figure 5 (a) Concept: all-optical wavelength conversion assisted by TriPleX micro-ring resonators, (b) experimental setup for 40 Gb/s continuous-

mode wavelength conversion

All-optical wavelength converter (AOWC) arrays. Each AOWC should be able to convert an optical packet to any of the different wavelengths supported locally by the router. Here we propose AOWCs implemented by cascading a single SOA and a periodic optical filter. Since SOAs and micro-ring resonators can be densely integrated in arrays, the SOA - TriPleXTM ROADM scheme can provide scalable and power efficient AOWCs.

Wavelength selective optical cross-connect (λ-OXC). The λ-OXC is responsible for switching optical packets from any input to any output depending on the wavelength assigned by the AOWCs. The λ-OXC can be implemented using cascaded TriPleXTM ROADMs that will “drop” each packet to a specific output. By tuning the ROADMs the λ-OXC can be fully reconfigured supporting dynamic changes in the network topology.

The SOA - TriPleXTM ROADM concept and its experimental demonstration are presented in the next sub-sections.

B. All-optical wavelength conversion using Si3N4-SiO2 micro-ring resonators

A key functionality for realizing high-speed photonic routing is the capability to wavelength convert packet-based traffic using high-speed, small footprint and energy efficient integrated components. A promising approach for implementing such wavelength converters is by exploiting the chirp induced through cross gain modulation of interacting pulsed and CW signals within a SOA [5]. By filtering specific spectral components of the wavelength-converted signal, slow SOA recovery

temporal components are filtered, effectively speeding-up the impulse response of the wavelength converter system.

Figure 5 shows the ring-assisted AOWC (RAWC) concept. The data packets P1 and P2 enter the SOA serially as pump signals. After reading the header of P1 the router controller drives DFB_1 that generates a packet-length CW signal at wavelength λ1. This signal is launched in the SOA as probe signal, synchronized with P1. Similarly, the subsequent data packet P2 is temporally aligned with a packet CW at wavelength λ2. The pump signal modulates the SOA gain and via Cross-gain modulation (XGM), this modulation is transferred to the probe signal. As such, an inverted (wavelength-converted) copy of the pump signal appears at the output of the SOA, The temporal profile of the converted signal shows chirping due to the refractive index modulation in the SOA with the leading edge of the probe signal being red-shifted and the trailing edge being blue-shifted. A micro-ring resonator ROADM is cascaded after the SOA with the drop-port spectrum depicted in figure 2. The probe signals are selected to coincide with the transmission peaks of the ROADM and by detuning the micro-ring resonators with respect to the CW wavelengths either blue or red shift chirp filtering may occur, leading to the acceleration of the effective operational speed of the AOWC. Depending on the ROADM detuning, the signal at the output can be either inverted or non-inverted [6]. In the case of inverted AOWC, a cascaded Delayed Interferometer (DI) is required to restore pulse polarity. The spectral response of the DI is detuned so that one of the spectral “dips” of the notch filter is superimposed with the optical carrier. As such, the excess CW signal which remains un-modulated by the XGM effect in the SOA is removed and the polarity of the signal is restored.

Figure 8. Experimental setup for the packet-mode, WDM-enabled RAWC.

Figure 6. (a) Concept: all-optical wavelength conversion assisted by TriPleX micro-ring resonators, (b) experimental setup for 40 Gb/s

continuous-mode wavelength conversion

C. Experimental evaluation

Initially we evaluated the performance of the RAWC by converting a continuous data stream on a single CW wavelength. Figure 5(b) shows the experimental setup. The 40 Gb/s transmitter consists of a CW DFB laser at 1556.55 nm, an Elecroabsorption Modulator (EAM) driven at 40GHz for pulse carving and a Ti:LiNbO3 modulator to modulate the 40GHz clock and provide a 27-1 PRBS. The 40 Gb/s data stream pulses were compressed to 3ps using a non-linear fiber compressor.

The RAWC comprises the SOA, the 2nd order TripleXTM ROADM and a DI used only in the case of inverted operation. The ROADM consists of the two coupled Si3N4–SiO2 TriPleXTM microring resonators that can be tuned independently by on-chip heaters. The Free Specrtal Range (FSR) of the device is 4 nm and the FWHM bandwidth is 0.6 nm. The DI is implemented using two polarization beam splitters (PBS) and standard polarization maintaining (PM) fiber. The DI provided a differential delay of 2 ps between TE and TM polarization components. The 2ps delay corresponds to 500 GHz FSR and is chosen to match the FSR of the ROADM and ensure that the DI spectral “dips” will attenuate only the optical carrier of the wavelength converted signal whereas the rest of the spectrum will remain unaffected. The local probe wavelength is provided by a CW DFB at 1562.75 nm. Finally the

receiver part consisted of a 40-to-10 Gb/s EAM-based demultiplexer and a 10 Gb/s error detector. Figure 6 illustrates the experimental results recorded with a 80 GHz digital communication analyzer. The eye diagram of the incoming data is depicted in figure 6a). Figure 6b) shows the signal directly at the output of the SOA indicating the device slow recovery time. Firstly, the ROADM is detuned by 0.1 nm selecting the lower signal wavelength (blue-shift chirp) and the inverted signal of figure 6c) is obtained. The eye diagram reveals

Figure 7. BER curves for RAWC of continuous data

that the ROADM filtering accelerates the operational speed of the system within the 40 Gb/s bit slot. The inverted signal at the output of the ROADM is launched in the PM DI and the pulse polarity is restored (figure 6e). Identical experimental setup was used for non-inverted wavelength conversion with the exception of the DI. By detuning the ROADM 0.3 nm with respect to the probe signal, the non-inverted eye diagram of figure 6d) was obtained. Figure Ff) and g) show the eye diagrams of the 10 Gb/s demultiplexed signals. Figure 7 shows the BER measurements. The inverted wavelength converted signal exhibits a power penalty of 0.84 dB whereas a power penalty of 1.5 dB is measured for the non-inverted signal. The higher penalty for the non-inverted operation is mainly due to the higher detuning required, which leads to higher loss and higher OSNR degradation.

The performance of the RAWC was also evaluated by wavelength converting a sequence of two 40 Gb/s time-domain multiplexed (TDM) optical packets. Each packet is converted onto a new wavelength, verifying the system capability to operate in a WDM environment. Figure 8 illustrates the experimental setup of the packet-mode RAWC. The two 40 Gb/s, TDM data packets of the same wavelength (1551.1 nm) enter the RAWC, temporally synchronized with two different CW packets. The CW packets are generated using two DFB lasers (1555 nm and 1559.16 nm), modulated at the packet rate in a Ti:LiNbO3 modulator. A mux-demux stage and an optical delay line are used to provide the 2-wavelength TDM CW packet stream. Due to the temporal synchronization of the packet stream, the first data packet is converted to 1555 nm and the second packet to 1559.1 nm. In the RAWC, the multi-wavelength operation is enabled by the periodic response of the integrated ROADM.

Figure 9 illustrates the experimental results for the WDM-enabled operation. Figures 9a) and (b) show the pulse trace and the eye diagram of the inverted wavelength converted packets at 1555 nm. The same results for the wavelength converted packet at 1559.16 nm are depicted in figures 9c) and d). The eye diagrams reveal identical wavelength conversion for the two consecutive packets at the two different wavelengths. Figure 9e) shows the corresponding BER curves. Error free operation was obtained with a 0.8 dB power penalty for the data packets at 1559.1 nm and 1 dB power penalty for the data packets at 1555 nm. The optical power requirements were 7 dBm for the CW and 3 dBm for the data packets. The RAWC requires approximately 1.5 W of electrical power to operate, which includes the SOA bias and TEC currents as well as driving requirements for the micro-ring resonator heaters

IV. CONCLUSION

Realization of photonic integrated systems-on-chip

using micro-photonic integration technologies is a promising path for developing transmission and switching hardware of next generation optical networks. Planar lightwave circuits, acting as photonic printed

circuit boards can increase the integration scale and the aggregate throughput of photonic devices. The silica-on-silicon low index contrast material is capable for hybrid integration of III-V active elements such as lasers and modulators enabling high-capacity transmission and

regeneration systems. The TriPleXTM high-index contrast material is capable for ultra-small and power efficient components suitable for photonic routing platforms. Here we have reviewed the recent advances in component fabrication and system demonstration using these two promising material systems.

ACKNOWLEDGMENT

This work is supported by the European Commission through projects ICT-APACHE (www.ict-apache.eu) and ICT-BOOM (www.ict-boom.eu) under the 7th Framework Programme, Information and Communication Technologies (ICT). CIP Technologies, LioniX, XiO Photonics and HHI are gratefully acknowledged for their continuing innovation in research & development and for fabricating and packaging the photonic integrated components reported here.

Figure 9. BER curves for RAWC of continuous data

REFERENCES

[1] G. Maxwell, “Hybrid Integration Technology for High Functionality Devices in Optical Communications”, in proc. OFC 2008, San Diego, U.S.A., March 2008, invited paper OWI3.

[2] H. N. Klein et al., “1.55µm Mach-Zehnder Modulators on InP for optical 40/80 Gbit/s transmission networks”, in proc. IPRM 2006, May 2006, New Jersey, USA, pp. 171-173, TuA2.4.

[3] (2008). [Online] Available: http://www.lionixbv.com/ [4] R. Heideman, et al., “Large-scale integrated optics using

TriPleX™ waveguide technology: from UV to IR”, in Proceedings of the SPIE Photonics Packaging, Integration, and Interconnects IX, 6-28 January 2009, San Jose, California, United States, Volume 7221 (2009)., pp. 72210R-72210R-15.

[5] Y. Liu, et al., “Error-Free All-Optical Wavelength Conversion at 160 Gb/s Using a Semiconductor Optical Amplifier and an Optical Bandpass Filter”, J. Lightw. Technol., Vol. 24, No. 1, pp. 230-236, Jan. 2006.

[6] J. Dong, X. Zhang, S. Fu, J. Xu, P. Sum and D. Huang, “Ultrafast All-Optical Signal Processing Based on Single Semiconductor Optical Amplifier and Optical Filtering”, IEEE J. Sel. Topics Quantum Electron., vol 14, No3, May/June 2008, pp. 770-778.

Dr Efstratios Kehayas obtained the B.Eng. degree from Southampton University, Electronics & Computer Science Department, the MSc degree from Imperial College London, Blackett Laboratory and the PhD degree from the National Technical University of Athens, School of Electrical & Computer Engineering. Dr Kehayas has authored and co-authored more than 50 scientific journal and conference publications in IEEE and OSA, including invited talks in major conferences. Dr. Kehayas is a chartered Electrical & Computer Engineer and as a contractor has performed research and development activities in the photonics ICT sector for SMEs and research centers within Athens, Greece. In 2006 he co-founded EXELITE Innovations Ltd, the first R&D photonics company with manufacturing capabilities in the Greek sector, and acted as the company R&D Director. Since 2008, he is a member of the executive board of the Greek Photonics Technology Platform. Dr Kehayas has a track record in the organization, authoring and management of research projects. Since 2004, he has authored and co-authored several successful Greek- and European- funded research projects within the 6th and 7th Framework Programme including FP6-CRAFT-MULTIWAVE, FP7-ICT-APACHE and FP7-ICT-EURO-FOS. Dr. Leontios Stampoulidis graduated from the Electronic & Electrical Engineering Department of the University of Patras. He obtained the PhD degree from the Photonics Communications Research Laboratory at the National Technical University of Athens. He was actively involved in a number of EU-funded research projects within FP6, including IST-LASAGNE, IST-MUFINS and IST-E-PHOTON/ONE, whereas he has authored and co-authored FP7 projects ICT-BOOM and ICT-PLATON respectively. In the period 2008-2009 he was involved in running EU project ICT-BOOM on behalf of the coordinating institute ICCS/NTUA. Leontios co-founded EXELITE Innovations, where he acted as the Director of Technology Solutions. Dr. Stampoulidis has published more than 50 scientific papers in scientific journals and conferences. He is a member of IEEE and OSA.

Paraskevas Bakopoulos obtained his Diploma of Electrical Engineering and Computer Science from the National Technical University of Athens with specialization in telecommunications, in 2003. He is a member of the Photonics Communications Research Laboratory of NTUA since September 2003. His research activities focus on multi-wavelength laser sources for telecommunications and Terahertz imaging applications and on the design and development of high-speed multi-format switching and transmission systems.

A Peer-to-Peer Game Model using Punishment Strategies

Chunzhi Wang1, Hongwei Chen1, 2, Ke Zhou1, Hui Xu1, Zhiwei Ye1 1School of Computer Science and Technology, Hubei University of Technology, Wuhan, China

2College of Computer Science & Technology, Huazhong University of Science and Technology, Wuhan, China Email: chw2001@sina.com, chw75@sohu.com

Abstract—Recent years, with the rapid development of P2P networks, the security problem of these networks has become increasingly obvious. According to the behavior of selfishness and betrayal for nodes in P2P networks, this paper analyzes and contrasts the benefits of these nodes, and presents the P2P game model with the penalty factor. The reasonable analysis and simulation by gambit prove that, addition of the penalty factor has certain constraints on the betray nodes and promotes the active cooperation of these nodes, thus improves the security status of P2P networks. Index Terms—P2P networks, penalty factor, game model, gambit

I. INTRODUCTION

Recent years, applications of P2P networks become wider, which include cooperative handling, file sharing and distributed storage etc. P2P networks have become the mainstream development model for future networks. In P2P networks, all nodes are equal, which means that they are both client-side and server-side. Thus these networks have no center, self-organization as well as scalability. However, the nodes in P2P networks behave more selfishness and chase their own networks to maximize revenue [1]. Judging from all that mentioned above, it will cause following issues.

(1) The problem of free-riding. For example, a test [2] obtains the following data, which is that 70% of nodes in the networks do not share any files, and nearly 50% of information search is from only 1% of nodes.

(2) The problem of tragedy of common. As non-exclusive public possessions, the resources in P2P networks are downloaded and used by a majority of uncontrolled nodes in these networks.

Trust exists in the social life, which means that someone believes the future behavior of others. It can also be used to resolve the safety problem in P2P networks. Furthermore, it can resolve the trust problem

between nodes. By the game theory hypothesis which is based on the node’s entirely rationality, it can make the nodes rational and grasp their mental state among the processes during which they interact with others. For instance, do they trust them; how the nodes will change in strategy selection when their interactive objects betray them. The paper tries to sum up the game processes between the nodes and adjust a number of factors in order to make the nodes tend to trust other nodes.

Some studies indicate that it can establish an effective confidence-building model [3] to resolve such security problems to some extent. Currently, there are some fruitful results about studies on constructing trust models for P2P networks, which are summarized as follows.

(1) Trust models based on Bayesian networks. This kind of models gives a node two trust degrees, derives its certainty factor according to its cooperation in the network and also based on Bayesian networks, thus then it cooperates with other nodes selectively with trust degree as its premise. The weakness of this model is that, the network model considered in this model is too rational, and the nodes are not classified, thus in this case, some “extreme” nodes such as malignant nodes are not further analyzed, lacking corresponding constraint management methods.

(2) Trust models based on PKI. This type of models is supervised by a few leading nodes with CA certificates, which maintain the stability of networks, manage the trust degree, and synchronously notify and penalty offending nodes. But this type of models is maintained by a few nodes, thus is weak in extensibility and may cause the problem of one-node-disability.

(3) Trust models based on Eigen Trust algorithm. This kind of models promotes network nodes deal with each interaction rationally by certainty factors. And the global certainty factor of a node relies on its neighbor nodes and satisfactory degrees of interactive nodes historically. It also has weaknesses, which are the lack of considering the cooperative cheating action of malignant nodes in the network, the problem of deciding the trust degree of new adding node, and the complexity problem.

(4) Trust models based on NICE. In this type of models, nodes interact and create a cookie for other side.

This work was supported by Natural Science Foundation of HubeiProvince of China (2009CDB100), and Foundation of Wuhan TwilightPlan Project (201050231084).

According to the cookie being positive or negative, other nodes can judge the quality of interacting nodes. However, this type of models still cannot improve the problem of cooperative cheating action of malignant nodes in the network, and control new adding malignant nodes.

(5) Trust models based on RG Trust algorithm. This kind of models is different from methods that centralize trust management or formation of global trust degree by neighbor nodes. This kind of models combines game theory and uses improving policies, which makes the cooperation that is not completed in one interaction realized in multiple repeated interactions. Thus nodes will not choose to betray for only one benefit. This kind of models has problems, which are it can not differentiate penalties for occasional mistakes of gracing nodes and repeating mistakes of malignant nodes, and it has difficulty in holding random probability P.

According to the problems exist in these trust models, this paper combines forecast and Punishment Strategies together, and uses the Markov chain to forecast future multiple time scales for current network running environments, and based on the forecasting results, adopts corresponding Punishment Strategies. In detail, it counts violation actions of network nodes, thus effectively reduces the penalty for occasional mistakes of gracing nodes and deals with new adding nodes equally. Finally it promotes network nodes tend to cooperate in a macro view and maintains the stability for the running of P2P networks.

II. THE PLIGHT OF THE TRUST GAME Trust game[7] is applied as a game theory into P2P

networks; it treats the node’s trust relation as a game. In a simple model of a one-time trust game, it can be supposed that there are two nodes 1 and 2; they are ready to interact and have two strategies to choose cooperation or betrayal. And their choice will not affect others, following the table 1. The model indicated the loss as V (V>0). The communication loss is needed for each node to pay in the interactive process. At the same time, these nodes enjoy the services from the networks. The income is assumed as U, and U>V. Nodes choose different strategies, their revenue will be changed, and the specific distribution of nodes’ income is shown in Table 1. It is easy to find that the strategy combination (betray, betray) with the node 1 and 2 is the unique Nash equilibrium[8]. Because any nodes change their strategy from betray to cooperation, this change will make the benefits of the nodes from 0 to –V, these two nodes can not build the trust of one-time game. And the two sides have plunged into “Prisoner’s Dilemma” [9]. This is called as the plight of the trust game.

TABLE I. TRUST GAME MODEL

Node 2

Node 1

trust betray

trust U-V,U-V -V ,U

betray U,-V 0 , 0

Prisoner’s Dilemma is used to explain between the best

personal income and the best overall income. In real life, many areas appear this contradiction such as battle of the advertisement business, the arms race of the politics, etc. and the “plight” [10] is also appeared in the behavior game of the nodes’ interaction in P2P networks.

After analyzing the current trust game models[11] which are proposed by some scholars at home and abroad. The development of trust game model can be summarized as follows.

Reference [12] proposes a management model of nodes’ credit in P2P networks by analysis the plight in implementation of individual rationality and collective rationality. This model reduces the time complexity of calculating credit and packet traffic, but the model doesn’t consider the nodes’ collusion in P2P networks, as well as how to treat the new network nodes differently. Reference [13] is also based on the Prisoner’s Dilemma, and it proposes a method to improve the trust relationship between network nodes by adjusting the related factors of the game revenue’s mechanism. It solves the trust plight between nodes from the macro level. But more practical issues in P2P networks are not considered, such as how to select the appropriate strategy for nodes of different types in the network, and how to encourage the nodes contribute more resources. The model built by the Reference [14] is based on game theory, with the view of maintaining fairness for nodes. It allocates more bandwidth and incomes for the nodes which contribute more resources. This model sets motivating factors which are the nodes’ income and contribution, in order to achieve optimal Pareto about the allocation of bandwidth resources, ultimately maximize social benefits. However, some problems exist in this model; for instance, the model can not treat different between the new nodes and other nodes in existence, and adapt the active state about more frequent entering and exiting by nodes. Reference [15] proposes an incentive scheme which is based on the quality of differential services, but this scheme doesn’t consider the nodes’ collusion problem in P2P networks, and it doesn’t discuss the contribution of the new nodes.

After researching the correlation model, it can be found that the existing and mature trust game models[16] in P2P networks usually gives nodes the trust degree, or regulates the parameters of game mechanism[17], and then the model from this point to inspire nodes being more rational, or more trusted choice. But it always takes

the appropriate adjustment[18], which aims at the appeared question and does not consider from the forecast point to anticipate the future network environment, and can not take a more active strategy in advance based on predicting network state. This paper just uses this method[19] to take timely punishment strategies for the nodes’ behavior in the network, and carry on the constraint, supervision, and maintenance. The proposed method prevents the status of networks moving in disorder, which leads the network environment being more secure and more trustful in advance.

III. CONSIDER THE PENALY FACTOR OF THE NODE’S GAME MODEL

Through analysis of the trust dilemma above, the model of P2P trust game is proposed, and it is based on punishment strategies. The main idea of this model is that if the nodes take the betrayal strategy when they are trusted by the object, the networks will punish it immediately. This model will make the node rational and choose cooperation rather than betrayal in the coming interactive strategy. After the virtuous circle of this strategy, the node can get rid of the dilemma better and improve the trust problem in P2P networks. Ultimately, the P2P networks are able to be in the state of more cooperation.

A. The payoff matrix of different types of nodes This paper collects several nodes’ interaction and

summarizes the node type as follows. • Trusted nodes (Nc): these nodes always actively

choose the cooperative strategies when they interact with any node, they rarely betray others, or make mistakes.

• Betrayal nodes (Nd): these nodes always want to achieve the biggest gains through the selection strategies. So they choose non-cooperation strategies constantly.

• Free nodes(Nr): these nodes randomly select the strategy of non-cooperation or cooperation, this paper assumes that the probability for free nodes to choose cooperation strategy is P (0<P<1), the non-cooperation is 1-P.

The gains of different nodes have been summarized after N times game as the following tables.

TABLE II. THE GAINS OF TRUSTED NODES WITH OTHERS

Trusted

N(U-V)

2N(U-V)

Betray

N(W-V)

N(U-W)

N(U-V)

Free choose

NP(U-V)+N(1-P)(W-V)

NP(U-V)+N(1-P)(U-W)

N(1+P)(U-V)

TABLE III. THE GAINS OF BETRAYAL NODES WITH OTHERS

Betrayal

N(U-W)

N(W-V)

N(U-V)

Betray

Free choose

NP(U-W)

NP(W-V)

NP(U-V)

TABLE IV. THE GAINS OF FREE NODES WITH OTHERS

Free nodes

NP(U-V)+

N(1-P)(U-W)

NP(U-V)+

N(1-P)(W-V)

N(1+P)(U-V)

Betray

NP(W-V)

NP(U-W)

NP(U-V)

Free choose

NP2(U-V)+

1/2NP(1-P)(U-V)

NP2(U-V)+

1/2NP(1-P)(U-V)

2NP2(U-V)+

NP(1-P)(U-V)

The probability in the t moments that is supposed for the chance when the node encounters the node which chooses the cooperative strategy is P1. So it can be supposed

NNNPNNP++

+=1 (1)

In the t moments, the probability of the node

encountering the betrayal node is P2. So it can be supposed

NNNNPNP

++−+

Combining with the table 3, it is easily to get the total

earnings of Nc, Nr , Nd -- T(Nc), T(Nr) and T(Nd) is

21 )()()( PVWPVUNT c −+−=

211 )()1)(()()( PPVWPPWUPPVUNT r −+−−+−= (3)

1)()( PWUNT d −=

A prerequisite as W>V is considered, by the theorem: the sufficient condition satisfies the Nash equilibrium steady state and makes the trusted nodes always get the biggest earnings:

)()()( drc NTNTNT >> (4)

Proof: By the equation (1) , VWU >> , it will get: 0)()2()()( 21 >−+−−=− PVWPVWUNTNT dc

0)()()()( 21 >−+−=− PPVWPPVWNTNT dr

0)1)(()1)(()()( 21 >−−+−−=− PPVWPPVWNTNT rc

So it can prove that )()()( drc NTNTNT >> .

B. Establishing the punishment game model Through analyzing the influence that punishment

strategies affect on the income of three types of nodes, this paper constructs the punishment game model of the network nodes in a macro view, expressing as follows.

TABLE V. THE GAME MODEL WHICH HAVE THE PENALTY FACTOR

Node 2

Node 1

trust betray

trust U-V, U-V -V+W, U-W

betray U-W,-V+W 0 , 0

As shown in table V, it can be found that when the

node has betrayed its interactive objects that select trust strategy, the node will be punished and make it pay the value of W, W is the penalty factor set in this paper, and the assumption is W>V, at the same time, the victim node will receive the same benefit as earnings of W. As a result, penalty factor changes the Nash equilibrium point of the payoff matrix. The point in this game is (cooperation, cooperation).

IV. THE DISCUSSION OF PENALTY FACTOR Now that the penalty factor is discussed and analyzed,

the purpose of advancing the penalty factor is to increase the positivity for the nodes’ choice of cooperative strategy, and beating fluke mind of betrayal nodes and guiding the non-cooperative nodes will incline to choose cooperation after repeatedly interacting with other nodes. At the same time, the penalty factor will ensure the stability of the entire P2P networks.

A. Deduction of penalty factor

In this paper, formula of the penalty factor W is:

3 2 11 ( ) [ (1 ) ]-

a k tW V D k V p

t tπ π π

= + + + − −∑ (5)

The last time when the nodes choose betrayal is ta, D (k) is the number of the nodes chooses non-cooperative strategy when their objects choose cooperative strategy. When D (k) =0 and W=V, from the time t0, when the nodes enter the networks, if the nodes always maintain the cooperation with others, it will not be punished. When the nodes have the fluke mind in the first time at t moment, it betrays others, then in the next moment, the nodes will be punished, and should pay for W. From the causes above, it proves further consequence that the more

trust nodes in the network; the less the nodes betray others; the smaller D(k) will be, the farther the time when the nodes betray others than the last time, the

smaller0

1tt −

will be, so that the penalty factor will be

smaller.

B. Markov chains forecast function of the model

This paper summarizes the converted trends of three state nodes through counting up the behavior of network nodes, shown as follows.

Figure 1. transition diagram of network nodes states

As shown in Figure 1, each node can only transfer to itself or to the adjacent state, and can not stride the transition, for instance, the trustful node can not directly transfer to betrayal node. This transferred trend of states follows the behavioral rule of nodes in the network, that is if the nodes have the fluke mind, through gradual betrayal, their probability of cooperation will decrease, and then their transferred trend of states will follow as: trustful → free → betrayal.

Markov chains[20] are used for forecasting, with the formulas which are commonly used as follows.

If the state space I of Markov chains }0:{ ≥nEn is finite set, and its transition probability matrix ijP satisfies

IjiPij ∈∀> ,,0 , then there exists a only probability

distribution ),...,,,( 321 nπππππ = in I , and this distribution makes Iji ∈, and nIjiPi j /1),:min( ≤∈ have:

1) Pππ =

nP 1lim π=∞→

The mechanism counts the nodes’ status in every other t times based on the forecasting methods of Markov chains, supposing in t time, the status of three types of nodes in the network is )3,2,1,(, =∈ nEnEE , the initial distribution of node status is

))0(),0(),0(()0( 321 ππππ = , the transition probability matrix of nodes status can be built according the condition of nodes’ state transition. It can be assumed

that the number of nodes is im in iE state, the count of the nodes which from iE state transfer to the jE state

is ijm . The transition probability of states ijP can be

obtained, when the network nodes in iE state transfer to

the jE state, furthermore, iijji mmP /= . Then the

transition matrix of three states nodes in network, P can be concluded:

11 12 13

21 22 23

31 32 33

p p pp p p p

⎡ ⎤⎢ ⎥= ⎢ ⎥⎢ ⎥⎣ ⎦ (7)

The prediction mechanism can get the transfer case of the network nodes’ future status at any time, which is based on the transfer matrix of varies states of the nodes. So supposing that in n time, the transfer matrix about any types of the network nodes is nN PPP ).0(][ = , and by

P⋅= ππ , the steady-state distribution vector in regard to the three types of the network nodes(trustful , free , betrayal) is ],,[ 321 ππππ = .According to the distribution vector on various types of the nodes, this paper concludes the status equation S combines with the network running as follows

123 )1( PPPPS −−+= (8)

C. The combination of Markov and punishment factor

The management mechanism is installed in P2P networks in order to directly control and constraint the behavior of the network nodes. It can be assumed that the nodes enter the network in the first time, at the time T=t0, combining with the punishment strategies which are designed for the network, the running rules of the management mechanism are as follows. Step 1 First, the nodes’ interaction history in t time is

counted, the probability distribution of the three future states of the network nodes can be predicted, is ],,[ '

'10 ππππ = , it can make

decision that whether choose the punishment strategies or not to node i, and based on the initial conditions of the penalty factor.

Step 2 According to the result of the state equation S, if S>0, it is shown that the network is tend to deterioration, then the management mechanism will take the punishment strategies in the next time. Or if S<0, it can be found that the network in optimum operation, then the management mechanism will enter into step 3.

Step 3 If S<0, the mechanisms will keep supervising the environment of network in the next time, and access to step 1. If the network environment continues to maintain a good running, the mechanism will only be engaged in a supervisory role. However, if the environment is more

deteriorative than the predicted result in the last time, the management mechanism will enter into step 2 in time, and takes the punishment strategies to the network nodes.

Step 4 After the management mechanism executed the punishment strategies, it needs to compare the predicted result in last moment with the network future state which is lately used the punishment strategies. If the network state has a good trend, it can be confirmed that this punishment strategy is an effective way for the network, otherwise, the mechanism needs to adjust the parameters of the strategies continually, or to use a more accurate way to predict and to control in advance.

V. THE SIMULATION AND IMPLEMENTATION

First, this paper makes a pursuant analysis for the game behavior of a single node in P2P networks, and combines the interactional situation of the nodes with the actual and running environment of the network; finally, it simulates the nodes’ game behaviors by Gambit. The trend of nodes’ behavior selection is compared, which is based on whether the mechanism is carried on, so parameters this paper discussed above are gave real values, in order to have more intuitive analysis of the impact of the penalty factor for the game result of each node and the strategies selection of the nodes.

It can be assumed that when the nodes choose cooperative behavior, they will get the benefit U=9, and they must pay for the loss of network communication, V=4, penalty factor W=6, Figure 2 is the game simulation chart of the P2P nodes which do not execute the management mechanism, however, Figure 3 is the result about the game model simulation after the mechanism is ran. Observing the following two figures, it is clearly found that the nodes will more incline to cooperate with others by restraint and supervision from management mechanism, rather than betray others by a fluke in order to gain a poor income. So it can get the conclusion that the management mechanism have a more positive impact on the behavior choosing of the nodes, it can effectively maintain stability of the network in detail.

Figure 2. The simulation curve of the nodes’ game which is not

restricted by the management mechanism

Figure 3. The simulation curve about the nodes’ game which is restricted by the management mechanism

Secondly, the thesis simulates the real condition of the P2P network as a whole, obtains the real-time data, and draws a discriminative statistics about the three types of the nodes, before and after the management mechanism is executed. It can count the distribution probability of network nodes by Markov chains, ultimately, the next n-times transferred circs of the nodes are simulated and forecasted by Matlab. The probability distribution about the three types of nodes in the network can be assumed, is ]0,0,1[0 =π , the transition matrix of the three types

nodes’ states is 1P , which can be obtained based on the real-time data.

This paper predicts that the next 20 times transferred circs of the nodes which are not restricted by the management mechanism, as is shown in Figure 4. Through the graph, it can clearly observe the future trend of three types of the nodes; various nodes are affected by a fluke due to the management mechanism does not supervise and restrain, it leads the nodes take more choice of betrayal. However, the vicious behavior is not got an availably control by punishment, in the end, the operation environment of the network will be deteriorated in the future.

Figure 4. The transferred diagram of the nodes’ states out of restriction

from the management mechanism.

Figure 5. The transferred diagram of the nodes’ states by restriction

from the management mechanism

Because of the prediction about the transferred trend of the nodes in the future, in t+1 time, this paper installs the management mechanism in the network, and then the mechanism punishes the vicious behavior of the nodes in time, the transferred matrix about the nodes is 2p , which will be obtained in the next time by statistics for the transferred situation of the nodes. Now, the mechanism predicts the nodes’ transferred situation in next 20 times once again, and Figure 5 shows the simulated result. Compared to Figure 4, it can be clearly observed that the management mechanism has a positive impact on the nodes’ transferred situation, as long as the mechanism takes a restrained hand for the nodes’ behavior, a slight change will be produced in the nodes’ transferred trend, furthermore, the major change will be engendered in the future trend of nodes’ behavior in the whole network. It also can be assumed that the existing of the mechanism plays a positive role in punishing the nodes’ betrayal and network’s stabilization, and it makes accessorial action about the trust between the nodes, and the choice of the cooperation.

This paper also notes that the changes of a single node’s behavior selection will play an impact on the network. So the management mechanism is necessary, and it is also needed the mechanism to track and predict for the status of the nodes, in order to prevent the network from paralyzation.

VI. CONCLUSION This paper begins with the security issues of the P2P

networks and the plight of the node trust game, concludes the deficiency of correlative game model through researches the various types of trust models. Ultimately, the Markov prediction game model which is based on the punishment strategies is proposed. This paper classifies the type of nodes, through comparing with the influence of the game model which execute punishment strategies or not, then sets management mechanism which summarizes the transfer matrix of various types of nodes, and predicts the future state of network development.

After reasonable arguments and corresponding simulation experiments, it is clearly confirmed that the management mechanism with predictive ability plays a maintainable and preventive role in the network state. And the punishment strategies that are adopted in time by this mechanism, possesses a constrained and punitive action for the betrayal of nodes. Finally, this mechanism plays certain significance for resolving the security problems in P2P networks in a both micro and macro level.

ACKNOWLEDGMENT We would like to thank the reviewers and editors for

their detailed and valuable comments.

REFERENCES [1] Hughes D, Coulson. G, and Walkerdine J, “Riding on

Gnutella Revisited: the Bell tolls,” IEEE Distributed Systems Online, vol. 6, pp. 1–18, Jun 2005.

[2] Hardin G, “The Tragedy of the Commons,” Science, vol. 162, pp. 1243-1248, March 1968.

[3] Wang Y, Vassileva J, “Bayesian Network-based Trust Model,” Proceedings of the IEEE/WIC International Conference on Web Intelligence(WI’03), Halifx, Canada, pp. 372-378, 2004.

[4] Dou Wen, Wang Huai ming, “A recommendation based peer-to-peer trust model,” Journal of Software, vol. 15, pp. 571-583, Apr 2004.

[5] Sepandar D, Kamvar, Mario T, Schlosser and Hector Garcia Molina, “The Eigen Trust Algorithm for Reputation Management in P2P Network”.

[6] Khambatti M, Dasgupta P, Ryu K D, “A role-based trust model for peer-to-peer communities and Dynamic Coalitions,” Proceeding of the Second IEEE International, Information Assurance Workshop, New York, IEEE Press, pp. 141-154, 2004.

[7] Wang Tao, Lu Xian liang, “A Novel Peer-to-Peer Incentive Mechanism Based on Game Theory,” vol. 15, pp. 201-203, Feb 2006.

[8] C. Buragohain, D. Agrawal, S. Suril, “A game theoretic framework for incentives in P2P systems,” In: Proc. 3rd Int’1 Conf. Peer-to-Peer Computing. Los Alamitos, CA: IEEE Computer Society Press, 2003.

[9] M. L Littman, “Value-function reinforcement learning in Markov games,” j Of Cognitive System Research, vol. 2, pp. 55-66, 2000.

[10] Myerson R B, “Game theory: analysis of conflict,” Massaehusetts: Harvard University Press, 1991.

[11] Aberer K, Despotovic Z, “Managing trust in a Peer-to-Peer information system,” International Conference on Information and Knowledge Management, New York, ACM, pp. 310-317, 2001.

[12] Liu Ye, Yang Peng, “Study of Mechanism of Trust Management to P2P Networks Based on the Repeated Game Theory,” Journal of Computer Research and Development, vol. 43, pp. 586-593, Apr 2006.

[13] Man Hong fang, Yang Rong rong and Liu Feng ming, “Research of trust evolutionary mechanism based on game theory in P2P networks,” Computer Application, vol. 27, pp. 2710-2717, Nov 2007.

[14] Xu Hai mei, Zheng Xiang quan, Qi Shou qing and Nie Xiao wen, “Novel incentive based on game theory in P2P networks,” Application Research of Computer, vol. 25, pp. 2787-2788, Sep 2008.

[15] Chen Zhi qi, Su De fu, “Incentives Model in P2P Network Based on Game Theory,” Computer Engineering, vol. 31, pp. 118-120, Aug 2005.

[16] Dai Zhan feng, We Qiao yan, Li Xiao biao, “Recommendation Trust Model Scheme for P2P Network Environment,” Journal of Beijing University of Posts and Telecommunications, vol. 32, pp. 79-72, Mar 2009.

[17] Li Gong, “Peer-to-Peer Networks in Action,” IEEE Internet Computing, pp. 40-42, May 2002.

[18] Gou Jing, Wu Guo xin, Li Xiang, “Research and Design for Trust Model in P2P Networks,” Computer Technology and Development, vol. 19, pp. 102-105, Mar 2009.

[19] Joseph D, Grunwald D, “Perfecting Using Markov Predictors,” IEEE Transactions on Computers, vol. 48, pp. 121-133, Feb 1999.

[20] Jin Shao hua, “A Limit Theorem for Markov Chains,” College Mathematics, vol. 20, pp. 64-67, Aug 2004.

Chunzhi Wang(1963-), femal, from Hubei province of China, Master's degree, Professor of Hubei University of Technology, Dean of School of Computer Science, interested in the security of network and computer network, Computer supported cooperative work. The Chairman of Wuhan of CCF Young Computer Scientists & Engineers Forum(2010).

Hongwei Chen (1975-), male, from Hubei Province, PHD, Associate Professor of Hubei University of Technology, interested in Peer-to-Peer, Grid Computing, Information Security, Mobile Agent.

Ke Zhou(1987-), master candidate of Hubei University of Technology, interested in Peer-to-Peer.

Hui Xu (1983-), PHD, Lecturer of Hubei University of Technology, interested in network and service management. Since 2006, she has been a certified computer system analyst in P.R. China. In July 2008, her biography was selected for inclusion in the 26th edition (2009) of the Marquis Who’s Who in the World, California, USA.

Zhiwei Ye (1978-), male, from Hubei Province, PHD, Associate Professor of Hubei University of Technology, interested in Computational Intelligence, Image Processing.

Reliable Resource Search in Scale Free Peer-to-Peer Network

Wei Song1, Wenbin Hu1, Zhengbing Hu2, Xi Zeng1 1 Computer School, Wuhan University, Wuhan, China

2 Department of Information Technology, Huazhong Normal University, Wuhan, China Email: songwei@whu.edu.cn, hwb77129@126.com, hzb@mail.hust.edu.cn,zx0071@163.com

Abstract—The resource distribution and peer links in P2P network have an obvious scale free character follows power law distribution. Using this inherent character of P2P network to design resource search strategy is great significant for improving the search efficiency and reducing the costs. We analyze the scale free character in P2P network, and propose the reliable random walk search algorithm which can achieve high and reliable resource search through transferring query messages based on P2P power law distribution. We design simulation experiments to evaluate the performance of our reliable random walk. The simulation experimental results show that the reliable random walk based on power law distribution is a scalable resource search algorithm with high and reliable search efficiency and low search costs. Index Terms—Peer-to-Peer, resource search, complex network, scale free, power law distribution

I. INTRODUCTION

With the rapid development of computer and network technology, computing model is in progress of significant alteration. As a new network service model, P2P (peer-to-peer) network changes the way people share resources over Internet. Each node is not only a client, but also a server in P2P network system, whose resources are not stored on centralized servers, but on every distributed node local. In an unstructured P2P networks, nodes often use flooding and various optimized strategies to search desired resources. How to search resource effectively and efficiently in the case of lower network cost is a key issue in the study of P2P network.

A large number of complex systems in nature, such as the Internet, transport system, the spread of disease, various social networks and so on, can be described by network. The earliest research on network is made by mathematician, whose basic theory is Graph Theory. Classical graph theory tends to use regular topology to model real network, resulted in the emergence of regular network model. In the mid-twentieth century, Erdos and Renyi proposed that the establishment of a network connection is random and unordered. Based on this standpoint, the so-called random network model [1] is established, which had been the theoretical basis of scientists’ study of real network for a long time. Until recent years, researchers discovered that large numbers of real networks are neither regular networks, nor random networks, but networks with different statistical

characteristics, which are called complex networks. P2P network, with features such as small world [2] and scale free [3], is a typical complex network. Related research [4] on Gnutella, the largest P2P application, found that 70% of Gnutella users rarely share resources and nearly 50% of the resources hits contribute by only 1% Gnutella users. This distribution of resources and node degree shows obvious scale-free feature. It is of great significance to analyze the distribution of P2P connections to design efficient resource search strategy, and control the amount of messages.

Using the power law distribution of P2P network to design efficient resource search is a new research field. Current resource search method in P2P network is designed to focus on the nodes, and refer to nodes’ visit history and information of the neighbor nodes. These design principles increase the information processing of large numbers of additional nodes, and it is also difficult to measure the search efficiency and query costs. Our paper proposes an efficient search strategy, reliable random walk, on the scale-free feature of P2P network. And simulation experiments are designed to verify the analysis of the scale-free feature of P2P network and evaluate the performance of reliable random walk resource search.

Our Paper is organized as follow. Section 2 reviews the current related work, and in section 3 we give the assessment methods of scale-free feature of P2P network in reliable random walk. A detailed description of the proposed resources search strategy, reliable random walk, is given in section 4. Section 5 is for simulation and the experimental results are analyzed. Finally, in Section 6 we give conclusions and prospects of future work.

II. RELATED WORKS

Currently, unstructured P2P network is widely used in P2P applications. Gnutella [5], which is a pioneer of P2P applications, uses flooding mechanism to discover shared resources in network. Flooding method is simple and easy to follow, however it results in too much search messages. Therefore, there have been many improved algorithms proposed to reduce the search costs caused by flooding method. Reference [6] was the first introduction of improved search algorithms for unstructured P2P network, such as Iterative Deepening, Directed BFS and Local Indices. Iterative Deepening determines whether to continue forwarding based on the results of each query,

which will increase the query delay due to the need to wait for each query results. In order to reduce the amount of query messages, Directed BFS selects a subset of neighbor nodes to forward search messages rather than forward query messages to all neighbors. However, it will increase the node processing cost and result in the aggregation of query requests to the excellent service nodes. In Local Indices, nodes store the information of resources in a radius neighbors. So, it only need make responses in partial TTL rounds to reduce search costs. But it requires nodes to store a large amount of resource information. Other resource search strategies of unstructured P2P network also include Adaptive Probabilistic Search (APS) [7], Random walk [8], PeerRank [9], Freenet [10], assist P2P search [11], Scalable Query Routing [12], etc.

A variety of resource search strategies in unstructured P2P network are basically designed based on nodes’ history behavior, which needs additional statistical information. It makes against the stability of the routing efficiency and is difficult to fit the dynamic changes of P2P network. If resource search strategy is designed based on the overall dynamic characteristics of P2P network, it would be helpful to improve the search efficiency and overall performance of P2P network.

Current related study on P2P network shows that P2P network has an obvious scale-free property. M. Ripeanu and others used Crawler to make statistical analysis on the Gnutella network and found that node-degree distribution obviously comply with power law distribution. When E. Adar and other people research the Gnutella network Free Riding [4] phenomenon, they also have found that 70% of Gnutella users do not share any resources. While 25% of the nodes in the network handled 98% of the resource search requests. It also fully demonstrates that resource distribution and the search processing have a clear scale-free feature follow power law distribution, which has been found in many other study of P2P network applications [6] [7] [8]. Taking advantage of the scale-free feature of the P2P network as optimizing tools to improve the resource searching efficiency in P2P network is becoming a new research direction. However, at present the field of making effective use of scale-free feature and other characteristics of complex network to achieve measurable, efficient and stable network resource search is still a blank, which is the main motivation of our paper.

Nima Sarshar and others designed percolation search method [14] in complex network environment based on its scale-free feature and the percolation theory, which was also extended [15] into P2P network applications. This percolation search query is efficient and stable with the advantage that nodes need not to store large amounts of additional information. However, the percolation search needs resource redundancy for communication and construction of the suitable distribution structure. At present there are some resource search strategy [12] [16] use scale-free feature of P2P network to optimize resource search, but they are just qualitative using scale-free character, whose search efficiency is not stable with

the dynamic changes of the peer and resource distribution. So it is also difficult to measure their efficiency.

We analyze the scale free distribution of P2P network, and design reliable random walk resource search strategy in unstructured P2P network environment based on the network inherent distribution character. Compared with similar improved random walk search algorithms [16] [17] and other unstructured P2P network optimizing algorithms, in reliable random walk no additional storage of statistical information is required for routing forwarding, and search efficiency is reliable and measurable. What is more, it also has high search efficiency, low network cost and easy to be extended on other network applications. Moreover, the reliable random walk algorithm also has a good reference value for other complex network researches, such as ad hoc, sensor networks, grid, etc.

III. SCALE-FREE PROPERTY ASSESSMENT OF P2P NETWORK

P2P network is a kind of complex network which is found with apparent scale-free properties in network connection by related research [13] [18]. That is, the network connection is re-tailed, and the majority of the nodes have only a few connected edges, but there are little nodes having a large number of connected edges. The scale-free property of complex networks is considered that the node degree is subject to the so-called power-law distribution. Namely, the proportion of nodes with the degree K in networks is showed in Formula (1), where A is the ratio constant and τ is the power-law distribution index.

P(k)=Ak-τ (1) The main content of our paper is to design an efficient,

stable and measurable resource search strategy utilizing scale-free property in P2P network. Accurate assessment of the network’s scale-free distribution is very important for our research. As P2P network is a highly dynamic one and nodes act of great autonomy. We use sample analytical tools to evaluate the P2P network power law distribution. We set up a scale-free assessment module over P2P network, which is consist of network connection acquisition sub-module and scale-free distribution analysis sub-module, which is responsible for statistics on the nodes’ connection distribution in P2P networks. The system architecture is showed in Figure 1.

Figure 1. Scale-free analysis for P2P network

Be different from the centralized network resource search similar to Napster, this scale-free analysis module in reliable random walk is responsible only for the connection status of nodes in networks and scale-free analysis as well, which is not involved in the resource search. Network node periodic visits the scale-free analysis module to update the local network connection situation, getting the analysis of overall scale-free distribution. The study of Gnutella [13] found that the network connection is in strict conformity with the power-law distribution of Formula 1 in initial P2P network applications. However, with the extension of network application protocols, nodes which have small network degree (less than a certain threshold K) no longer meet this power-law distribution, while the nodes with larger network connectivity (greater than the threshold K) still meet it well. So, the presentation of the power-law distribution in Formula 2 better reflects the real scale-free distribution in P2P network. And it is also the basis of scale-free analysis in reliable random walk. Since the search requests obtain responses mainly in those highly connected nodes through the P2P resource searching process. The distribution of the highly connected nodes can still be described and analyzed through the power-law distribution are the main concern in reliable random walk.

P(k)=Ak-τ, τ>0, k≥K (2) The scale-free network analysis module may calculate

the proportion P(k) that nodes of different degree get through the network connection information reported by nodes. And then uses the function approximation method to determine the scale-free parameter A and τ as well as the scale-free phenomenon inflection point K based on Formula 3. Assessment methods are described below. First, assign u=lny, v=lnx, and then transform Formula 3 to Formula 4 as a straight line form. Afterwards, take the advantage of the statistical data in network connection obtained to determine K, the inflection point of scale-free phenomenon. And next, use the least-squares fitting a straight line to determine the power-law distribution parameters A and τ in range of the scale-free distribution interval.

Ax-τ, τ>0, x≥K (3) u=lnA-τv (4)

In the scale-free evaluation of reliable random work, a so-called lazy synchronization mechanism is in use to collect the distribution of nodes connectivity. As the scale-free statistical analysis module does not take the initiative to obtain the node-degree information but by the network nodes periodic report the local network connection status. While a new node joins the P2P network, it would have access to the statistical analysis module for the current network scale-free distribution and establish a stable network connection, then notify the module its connectivity degree. Using this synchronization mode can greatly reduce the amount of network message caused by simultaneous network connection status. Moreover, statistical analysis modules can be relatively real-time obtain network connection status.

This method ensures the accuracy of the scale-free analysis, thus the follow-up resource search strategy design is reasonable. Finally, the simulation experiments will be designed to evaluate the correctness of scale-free assessment strategies in reliable random work.

IV. RESOURCE SEARCH IN P2P NETWORK BASED ON SCALE-FREE PROPERTY

Reliable random walk search strategy which is based on network scale-free distribution, allocates resource hitting probability to each random walk search message to achieve the scalability and reliability of resource search. The resource search messages of reliable random walk are generated and distributed entirely based on the P2P network power law distribution, which can effectively improve the resource search efficiency and minimize network cost for unstructured P2P network.

In reliable random walk, nodes visit scale-free analysis module periodically to obtain scale-free distribution of the network, including the scale-free distribution parameters A, τ and the scale-free phenomenon inflection point K. Resource search strategy is designed based on the scale-free distribution of the network. Figure 2 shows the structure of reliable random walk’s resource search query message. When resource search request is initiated by the node, the query condition Query is generated firstly. Related research [4],[13] on P2P network application found that resource search request in P2P network is responded mostly in the nodes with high node-degree. Therefore, in reliable random walk resource search, Degreehigh is set according to the network distribution. The central nodes are defined as nodes whose connectivity is not less than Degreehigh. When a search request message reaches a central node, it can be satisfied since the central nodes contain most of the resources information in the local network. The search request should not continue forward, which minimize the amount of network messages while the resource search efficiency does not be reduced. In a search request message, q represents the probability of hitting the central nodes undertaken by this message, which is used to quantify the search efficiency in reliable random walk. This probability is assigned to each resource search message to ensure the searching efficiency. The assignment of hitting probability and the distribution of search message will be further analyzed in next section. TTL represents the maximum radius of search request spreading.

Figure 2. Resources search query message of reliable random walk

Suppose the parameters of node P in P2P network obtains scale-free distribution is A and τ, scale-free inflection point is K. In reliable random walk, the central node, namely Degreehigh, is determined by an adaptive strategy. Degreehigh=DegreeMAX/2 in initialization and it can be updated by nodes in range of [K, DegreeMAX] according to the query situation. If a node wants to get information more, it can assign a bigger Degreehigh, and

vice-versa. When a query source node P generates a query condition Query and expects query message to hit the central node with probability QALL, it needs to calculate Qhigh, the proportion of central nodes according to the network distribution. Formula 5 shows the proportion of central nodes in P2P network, making use of scale-free distribution property of P2P network showed in Formula 2.

−=maxDegree

Degreekhigh

AkQ τ (5)

How can ensure to hit the central nodes with expected probability QALL, it is the core issue of reliable random walk to forward resource search request depends on the power law distribution of P2P network. Suppose that the largest transmission radius of query message is TTL, and the probability of search request hit the central nodes in each TTL rounds. Formula 6 is necessary in order to meet the overall resource search hit rate. Suppose that the hit probability of each TTL round resources search is roughly equal, then it can be considered that the central node hit rate of each TTL round Q meets Formula 7. Source node of the search request P calculate how many search request is needed based on the central node hit rate of each TTL round Q. Suppose the network connections of P is k, then in order to meet the hit rate Q, the amount of search request r is shown as Formula 8 and central nodes hit rate that every search request need to take on q is shown as Formula 9. Query request messages carry their own hit rate q to forward, which is keeping fulfilling in the following forwarding TTL to achieve the reliability and metrizability of overall resource search in reliable random walk.

1-(1-Q1)(1-Q2)…(1-QTTL)≥QALL (6)

TTLALLQQ −−≥ 11 (7)

)1ln()1ln()1(1

rhigh Q

QrQQ−−

≥⇒≥−− (8)

r Qq −−= 11 (9) In reliable random walk, the total query hit is break

down into each random walk search request. Unfortunately, a sufficient number of random walk (k<r) is not allowed according to the connectivity degree k of some nodes. It would undermine the stability and reliability of the overall resource search efficiency of reliable random walk. The following context of the specific query process is to analyze how to ensure the system query efficiency in this case.

Resource search process is shown in Figure 3, in which a search request source node P expects that the search request hits the central node in a probability of 90%. Firstly, calculate the proportion of the central node by Formula 5 Qhigh=5.26%(DegreeMAX=10), and set the maximum search radius TTL = 6, then in order to achieve the total search hitting rate as 90%, for each TTL round the probability of hitting the central node needs to be over 31.87%(Q>31.87%). When a node P delivers a search request, in order to satisfy the first-round search hitting probability the numbers of random walk r should be not less than 7.10 as computed by Formula 8 (r>7.10).

Unfortunately the network connection degree of the node P can not meet the requirement of publishing 8 random walks. In reliable random walk, when the search requests published by a node can not meet the requirements of the central node hitting probability, the node forwards or transits the query messages by its own network connection number k, during which each random walk carries the hitting probability of meeting the expectations but not make the hitting probability in each round as the allocation criterion. The search hitting probability will be submitted completed in the follow-up TTL round to ensure the overall source searching efficiency.

4%,903,5.2,65.0==

highALL DegreeQKA τ

Figure 3. Reliable search based on power law distribution

As shown in Figure 3, node P releases 3 random walks, using Formula 10 to calculate the hitting central node rate of each message taking q = 53.58%, putting q in each of the random walk. Based on the local network scale-free distributed property, nodes which received query messages use a similar strategy to calculate the number of random walk to be transmitted. In order to complete the hitting rate of each search message. Supposing the neighbor nodes obtain the same scale-free distribution as P, then in the remaining TTL-1 round, the hitting rate of each message transmitting round is expected to reach , and

the number of random walk needs to be transmitted is .In Figure 3 node A2 can transmits 3

query requests to meet the hitting rate of last round of random walk query. And each random walk possesses

hitting rate of the central nodes. The nodes A1 and A3 in Figure 3 can not transmit three random walks, therefore nodes transmits the random walk with their best capacity, and calculate the query hit rate of every random walk should take by Formula 11. The following query process is similar.

Our design is also reasonable, the amount of transmitted messages is calculated by Formula 8, while in Formula 8, the central node hitting rate Q of each round is supposed are the same in every TTL rounds. However, when the TTL is small, the quantity of query message volume is small, and the total node degree is low, so it’s impossible to obtain the same hitting rate of central nodes as the following rounds. For this reason, in the beginning of reliable random walk search process, search request information always determines central node hitting rate of every query round according to the expected hit rate QALL.

So, nodes try their best to forward search request information. In the following TTL searching process, the amount of query messages is determined to minimize network messages according to hitting rates of each round. Therefore, reliable random walk is still efficient to the whole search as long as it can achieve the hitting rate of every search message. In the simulation experiment we will analyze the situations which evaluate the resource search message hitting the central nodes in the reliable random walk.

rALLQq −−= 11' (10)

r qq −−= 11' (11)

In reliable random walk resource searching, when TTL=0 or the central node is hitting, the query message stops forwarding. The former stop condition is for the reason of controlling the query radius and the later one is in the consideration of reducing the amount of query messages under the precondition of ensuring the query efficiency. If the probability of a central node appears is Qhigh=5.26%, then the probability that a query message in the query range of TTL=6 hits two or more central nodes is

. Such a probability is very low, which is the main reason why the query message stops forwarding when it has hit a central node.

The reliable random walk resource searching makes the searching hit and random walk forwarding by using the scale-free characteristics of P2P network. When the connectivity degree of a node cannot ensure the query messages that have been sent reach the excepted hitting probability, the messages will be forward in the utmost. Otherwise, the messages will be forward in a number that can ensure the expected search hitting probability.

From the analysis of the retrieval process, it can be found that during the spreading of the query message, the query hitting probability of every random walk drops rapidly. In the following forwarding, no more query messages are needed. Such a resource searching strategy is conducive to rapid discovery of the central node and then response to the searching request. Besides, it can also control the amount of the messages in the network. Such a method is conducive to the stability of the unstructured P2P network resource search. In the following experiment section, related experiments are designed to evaluate the performance of reliable random walk resources searching.

V. SIMULATION EXPERIMENTS AND RESULTS ANALYSIS

The following simulation experiments are designed to simulate reliable random walk resource search strategy and evaluate its performance. Statistic of its scale-free analysis and efficiency in resource search has been made, and we compare reliable random walk with similar resources search algorithm to analyze the simulation results.

A. Experiment Setting Reliable random walk is a measurable and efficient

resource search algorithm base on unstructured P2P network’s scale-free distributed properties. In the simulation experiments, using PLOD algorithm [19] we get a P2P network scale-free topology with τ=2.5-3.0, A=0.60-0.75 to simulate the actual P2P network application environment, and design simulation network scale-free turning point K = 3, that is nodes with k≥3 meet scale-free distribution. The network size of simulation experiments is 1000-5000, and the maximum node connectivity degree DegreeMAX=10-20. Experiment 4.2 and 4.3 test and evaluate the scale-free analysis accuracy and shooting proportion of central nodes in the reliable random walk, which is implemented by JAVA. Experiment 4.4 tests the resource search strategy of reliable random walk, besides, compares the reliable random walk with the Gnutella and k-random walk search strategy, which are built by PeerSim [20]. Tests are in the same experimental environment and network parameters.

All simulation experiments are performed on a single PC, with configuration of CPU P4 2.8GHz, Memory 1GB, and Windows XP operating system. Table 1 shows the basic parameter settings of simulation experiments.

TABLE I. PARAMETER AND SETTINGS IN THE SIMULATIONS

Parameter meaning Value N Network size 1000, 5000

A Proportional coefficient of scale-free distribution 0.60, 0.75

τ Index of scale-free distribution 2.5, 3.0

K Inflection point of scale-free phase change 3

TTL Maximum forwarding hops of query messages 6

DegreeMAX Maximum connectivity degree of nodes 10, 20

QALL Expected hitting proportion of central nodes 80%, 90%

B. Scale-free analysis of P2P network Reliable random walk resources search expands

entirely based on the scale-free distribution properties of P2P networks. The accuracy of the network scale-free analysis is very important to the resource search, so we firstly design the simulation test to detect the accuracy of scale-free analysis in the reliable random walk. Maintain the network size as N, and all nodes join in and leave the network randomly every 300 seconds, and the scale-free analysis module reports local network connection information every 600 seconds. In practical P2P networks, it is difficult to collect network connection information of all nodes, so in the simulation experiments, the scale-free analysis module just selects only 20% network connection information of the latest nodes for network scale-free property analysis. Simulation experiment tested in different scale-free distribution and network scope to compare the difference on the network practical node degree distribution and node degree distribution in the P2P network scale-free module with scale-free analysis. Experiments last 5

hours, we collect the data of current node distribution in the network and evaluation data of scale-free module analysis every 5 minutes, then average and compare both of the results. The comparison is shown in Figure 4.

k=3 k=4 k=5 k=6 k=7 k=8 k=9 k=100

Peer degree

N=1000, real networkN=1000, evaluation networkN=5000, real networkN=5000, evaluation network

k=3 k=4 k=5 k=6 k=7 k=8 k=9 k=100

Peer degree

(a)τ=2.5, A=0.60 (b)τ=2.5, A=0.75

k=3 k=4 k=5 k=6 k=7 k=8 k=9 k=100

Peer degree

k=3 k=4 k=5 k=6 k=7 k=8 k=9 k=100

Peer degree

(c)τ=3.0, A=0.60 (d)τ=3.0, A=0.75

Figure 4. Analyses of Peer-to-Peer network power law distribution

In the reliable random walk, resource search shooting focuses on the central nodes with a high degree of network connectivity, therefore, scale-free analysis module makes use of the collected node data of network connection to analyze scale-free phase transition point K, as well as scale-free parameters A and τ, to determine the network scale-free distribution, and predict the distribution of nodes with a high network connection level (k>K). The experiment results show that by evaluating connection distribution of P2P network with scale-free analysis module and computing the distributed proportion of high connectivity nodes in the network according to the evaluation result, we can make the predicted distribution almost the same as the actual node distribution in the network. It can be found in the experimental results that the design of P2P network scale-free analysis strategy is effective and nodes can correctly understand other nodes distribution in the network with the help of analysis results, so as to ensure the correctness and effectiveness of following resource search strategy based on scale-free properties.

C. Query hitting probability of central nodes Reliable random walk resource search is designed

based on the reliable hit central node, so it is of great importance for the stability and scalability of reliable random walk's efficiency in searching resources. We care about whether reliable random walk resource search request message can hit the central node of the P2P network stably and reliably according to expected probability. We design simulation experiments to study the issue under different network conditions. The largest connectivity degree of the network nodes DegreeMAX=10, and the simulation experiment determine the central node in accordance with Degreehigh= DegreeMAX /2=5. The network size is maintained at N=5000. Every 60 seconds network nodes make a search request to the network with the expected probability QALL, TTL = 6, and obtain scale-

free distribution conditions every 600 seconds by accessing the scale-free analysis module. After 5 hours, we make a statistics of network query request's hitting probability of the central node according to the network run-time. The simulation experiment results are shown in Figure 5.

0.5 1 1.5 2 2.5 3 3.5 4 4.5 580

Time(hour)

abilit

expected hitting probability Q=90%expected hitting probability Q=80%

0.5 1 1.5 2 2.5 3 3.5 4 4.5 580

Time(hour)

abilit

(a)τ=2.5, A=0.60 (b)τ=2.5, A=0.75

0.5 1 1.5 2 2.5 3 3.5 4 4.5 570

Time(hour)

abilit

0.5 1 1.5 2 2.5 3 3.5 4 4.5 5

Time(hour)

abilit

(c)τ=3.0, A=0.60 (d)τ=3.0, A=0.75 Figure 5. Rate of hitting high degree nodes

Experimental data shows that reliable random walk resources search strategy can ensure that the search requests can hit the network central nodes within the expected query hit probability QALL in the query scope TTL limited. The hit probability of central node is slightly lower than expectations when the proportion of the network central node is very low(in Figure 5c situation, the ratio of the central node is 1.19%). Analyzing the results, when we query the nodes whose continuous hit rate in querying messages is extraordinary low in this case, the query message cannot be spread effectively, leading to a result that the central node cannot be hit effectively. In this case reliable random walk is almost equivalent with flooding messages. Nodes which have received query messages will transmit them with maximum capability, and their failure to hit the central node is result from network topology. Therefore, reliable random walk is a kind of adaptive resource search strategy, the nodes can set query request according to their own needs and the network scale-free distribution so as to achieve an optimal resource search. Next we will compare reliable random walk with other similar P2P network resources search strategies and make a further analysis about the search performance of reliable random walk.

D. Search efficiency of Reliable random walk Reliable random walk is a search strategy for

unstructured P2P network resources based on networks with scale-free characteristic, of which query efficiency can be measured and assured. In order to test the resources search efficiency of reliable random walk, we design some simulation experiments to make a comparison with other resources search strategy such as Gnutella, k-random walk. We use PeerSim [20] to build Gnutella, k-random walk, and reliable random walk

network protocol, and we validate the efficiency of their routing in the same experimental environment. The network topology of the simulation experiments is a Power Law distribution with different distribution conditions. Its scale-free phase-change point K=4, namely, the nodes whose k≥4 meet the scale-free distribution. The network size N=5000 and DegreeMAX=20. The resources each node stored accord with the Zipf distribution that α=1.2, DocumentMAX=200, and the number of search request each nodes initiated accord with the Zipf distribution that α=1.0, QueryMAX=100.

Firstly, simulation experiments are designed to compare the recall of three kinds of resources search strategy - Gnutella (flooding), k-random walk (k-RW for short, k = 1, 2, 3), and reliable random walk (RRW for short) under the condition of different P2P networks scale-free distribution. Reliable random walk node search network resources and make statistics of network average recall rate in accordance with TTL = 7 and expected probability of 90%. The results of experiment are shown in Figure 6. The experimental results show that recall rate of reliable random walk can be basically guaranteed in about 80% when TTL> 5, and it can be nearly 90% when TTL> 6. Reliable random walk is designed based on the distribution of P2P network resources, the experiment data show that reliable random walk can achieve a equivalent recall rate with the flooding strategy of Gnutella as the network changes in the distribution of scale-free, and it shows better adaptability for the changes of peer networks scale-free distribution compared with k-random walk. Meanwhile reliable random walk gets a higher recall rate compared with k-random walk when the TTL is small, which is conducive to improving the k-random query delay.

2 3 4 5 6 70

Gnutella1-RW2-RW3-RWRRW

2 3 4 5 6 70

Gnutella1-RW2-RW3-RWRRW

(a)τ=3.0, A=0.75 (b) τ=2.5, A=0.60

Figure 6. Recall of Gnutella, k-random walk and reliable random walk

Because the amount of searching messages is also a quality index to focus on in resource search, the simulation experiments was designed to compare the amount of query messages among different networks with the condition of scale-free distribution. As the messages of k-random walk is constant with the change of TTL, only the comparison of the query message amount between Gnutella and reliable random walk is necessary. In the simulation experiment reliable random walk node searches network resources in accordance with TTL = 7, and the expected probability is 90%, then we make statistics of query messages amount of each TTL message forwarded round. The experimental results are showed in Figure 7. From the experimental data it can be found significantly that as the TTL query messages increasing the amount of query message of reliable random walk is

far less than the amount of Gnutella’s query message, while the recall of reliable random walk was almost equal with Gnutella. It is also found in the simulation experiment that as the TTL increases the query message amount of reliable random walk is not increasing sharply. A in-depth analysis revealed that the query message is not forwarded after the central node is hit in query process, which greatly reduces the growing query message later, and because the central node stores a large number of network resource information, queries can be correctly responded, which ensures that the query efficiency will not be affected.

2 3 4 5 6 70

reliable random walkGnutella

2 3 4 5 6 70

reliable random walkGnutella

(a)τ=3.0, A=0.75 (b) τ=2.5, A=0.60

Figure 7. Query messages of Gnutella and reliable random walk

VI. SUMMARY AND FUTURE WORK

This paper presents reliable random walk, a resource search algorithm for unstructured P2P network. It is based on the scale free character which is an inherent character of P2P network. The main advanteges of reliable random walk is that its searching efficiency is reliable and metrizable. Reliable random walk predicts the proportion of the central nodes, uses query messages hitting the central node to quantify the resource search efficiency of the P2P network. Each query message in the network brings the quantitative hitting probability, which realizes the metrizability and high reliability of P2P network resource search.

Compared to other similar P2P network resources search algorithms, reliable random walk has lower maintenance costs, need not to store any additional supporting information, and does not require statistical analysis on the information of the history search requests. In addition, reliable random walk is easy followed in unstructured P2P network and easy to implement. During the experiment we found that the deterioration of local network condition, such as low local network connectivity, will cause some reliable random walk query failure. Therefore, in future research work, we'll consider the introduction of local scale-free feature analysis of the node to enable node’s control of local P2P network’s distribution. In this way query messages can be adjusted adaptively according to the distribution of local network, which will make the network query scalable.

ACKNOWLEDGMENT

This research is partially supported by Natural Science Foundation of China under Grant No. 70901060, China Postdoctoral Science Foundation under Grant No. 20100471145, the Fundamental Research Funds for the Central Universities under Grand No. 6082024, National

the Natural Science Foundation of Hubei Province under Grant No. 2009CDB304.

REFERENCES

[1] P.Erdos, A.Renyi. On the Evolution of Random Graphs. In: Publications of the Mathematical Institute of the Hungarian Academy of Sciences, 1960, 5:17-61

[2] D. Watts, S. H. Strogatz. Collective Dynamics of Small World Networks. Nature, 1998, 393: 440-442

[3] Reka Albert, Albert-laszlo Barabasi. Statistical Mechanics of Complex Networks. Reviews of Modern Physics, 2002, 74: 47-97

[4] Eytan Adar, Bernardo A. Huberman. Free Riding on Gnutella. Xerox PARC, 2000

[5] Gnutella. http://www.gnutella.com [6] Beverly Yang, Hector Garcia-Molina. Improving Search in

Peer-to-Peer Networks. In: Proc. of the 22nd International Conference on Distributed Computing Systems (ICDCS), 2002, pp. 5-14.

[7] Dimitrios Tsoumakos, Nick Roussopoulos. Adaptive Probabilistic Search for Peer-to-Peer Networks. In: Proc. of the 3rd International Conference on Peer-to-Peer Computing (P2P), IEEE, 2003, pp. 102-109.

[8] Qin Lv, Pei Cao, Edith Cohen. Search and Replication in Unstructured Peer-to-Peer Networks. In: Proc. of the 16th ACM International Conference on Supercomputing (ICS), New York, ACM Press, 2002, pp. 84-95.

[9] Feng Guo-Fu, Mao Ying-Chi, Lu Sang-Lu, Chen Dao-Xu. PeerRank: A Strategy for Resource Discovery in Unstructured P2P Systems. Journal of Software, 2006, 17(5): 1098-1106.

[10]Ian Clark, Oskar Sandberg, Brandon Wiley, Theodore W. Hong. Freenet: a Distributed Anonymous Information Storage and Retrieval System. In: Proc. of the Workshop on Design Issues in Anonymous and Unobservability. Heidelberg: Springer-Verlag, 2000, pp. 311-320.

[11]Rongmei Zhang, Y. Charlie Hu. Assisted Peer-to-Peer Search with Partial Indexing. In: 24th Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), New York: IEEE Press, 2005, 1514-1525.

[12]Abhishek Kumar, Jun Xu, Ellen W. Zegura. Efficient and Scalable Query Routing for Unstructured Peer-to-Peer Networks. In: 24th Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), New York: IEEE Press, 2005, 1162-1173.

[13]Matei Ripeanu, Ian Foster, Adriana Iamnitchi. Mapping the Gnutella Network: Properties of Large-Scale Peer-to-Peer Systems and Implications for System Design. IEEE Internet Computing Journal, 2002, 6(1): 50-57

[14]Nima Sarshar, Oscar Boykin, Vwani Roychowdhury. Scalable Percolation Search on Complex Networks. Theoretical Computer Science. 2006, 355(1): 48-64.

[15]Nima Sarshar, P. Oscar Boykin, Vwani P. Roychowdhury. Percolation Search in Power Law Networks: Making Unstructured Peer-to-Peer Networks Scalable. In: Proc. of the 4th International Conference on Peer-to-Peer Computing(P2P), IEEE, 2004, pp. 2-9.

[16]Nabhendra Bisnik, Alhussein A. Abouzeid. Optimizing Random Walk Search Algorithms in P2P Networks. The International Journal of Computer and Telecommunications Networking. 2007 51(6): 1499-1544.

[17]Ming Zhong, Kai Shen. Popularity-Biased Random Walks for Peer-to-Peer Search under the Square-Root Principle. In: Proc. of the 5th International Workshop on Peer-to-Peer Systems (IPTPS), Santa Barbara, CA, 2006.

[18]Stefan Saroiu, P. Krishna Gummadi, Steven D. Gribble. A Measurement Study of Peer-to-Peer File Sharing Systems. In: Proc. of the Multimedia Computing and Networking (MMCN), SPIE/ACM, San Jose, CA, USA, 2002, pp. 18-25

[19]Palmer CR, Steffan JG. Generating Network Topologies that Obey Power Law. In: Proc. of the Global Internet Symposium (Globecom), San Francisco, IEEE, 2000, pp. 434-438.

[20]PeerSim. http://peersim.sourceforge.net/

Wei Song received the BS degree in Mechanical Science and Engineering from Huazhong University of Science and Technology, China in 2001, and Ph.D. degree in Computer Science from Huazhong University in 2008. He is currently a lecture and post-doctor in the School of Computer at Wuhan University. His main research interests

include peer-to-peer network, distributed system, and distributed system security.

Wenbin Hu Associate professor worked in Wuhan

University since 2006. He got his Ph. D degree in 2004. His main research interests include intelligent simulation and optimization, multi-agent system and swarm intelligent algorithm.

Zhengbing Hu received B.E., M.E. and Ph.D. degree in

National Technical University of Ukraine. He is currently a lecture in Huazhong Normal University. His current research interests include network security, intrusion detection system, artificial immune system, data mining etc.

Xi Zeng received his bachelor’s degree in Computer School

from Wuhan University in 2010. Now he is pursuing his M.S. degree in Department of Computer Science and Engineering at the Chinese University of Hong Kong, Hong Kong, China. His main research interests include distributed systems, peer-to-peer network.

Tree Routing Protocol with Location-based Uniformly Clustering Strategy in WSNs

Gengsheng Zheng

School of Computer Science & Engineering, Wuhan Institute of Technology, China Email: zhenggengsheng@sina.com

Zhengbing Hu

Department of Information Technology, Huazhong Normal University, China Email: kievpastor@yahoo.com

Abstract—In WSNs, energy efficiency and low latency are considered as two key issues in designing routing protocol. This paper introduces a two layer hierarchical tree routing protocol with location-based uniformly clustering strategy in WSNs (TRPLUCS), which can gives a good compromise between energy consumption and delay. First, TRPLUCS makes balanced clustering based on geographic regions. That is to say, all the clusters are uniformly distributed in WSNs with approximately equal number of nodes in each cluster. Second, TRPLUCS computes a spanning tree inside the cluster. Each node in cluster transmits data to its cluster-head along the spanning tree. Third, TRPLUCS constructs an optimal tree among all the cluster-heads. The data packets are then routed to the sink along the optimal tree. Simulation results show that TRPLUCS performs better than PEGASIS and LEACH in terms of energy*delay cost metric. Index Terms—wireless sensor networks, energy, delay, cluster, hierarchical tree

I. INTRODUCTION The requirement for designing routing protocol in

WSNs is different from traditional wireless network. The main constraint of WSNs is their very limited battery energy, which has great influence on the lifetime and the quality of the network ([1], [2], [3]). On the other hand, transmission time is also another important factor to a routing protocol in the real-time environment ([4], [5]). For these reasons, the protocols running on sensor networks must consume the energy efficiently while keep a low latency.

Most of the routing protocols are either data-centric or hierarchical or location-based ([6], [7], [8], [9], [10], [11]). Among them, hierarchical routing protocols provide better solutions than other schemes in the energy awareness and real time application. In hierarchical routing protocol, one node acts as the cluster-head to collect the data in each cluster. Data communication in different clusters can process simultaneous. Among the hierarchical protocol Low-Energy Adaptive Clustering Hierarchy (LEACH) ([12], [13]), Power Efficient Gathering in Sensor Information Systems (PEGASIS) [14], Chain Oriented Sensor Network for Efficient Data

Collection (COSEN) [15], and Base-Station Controlled Dynamic Clustering Protocol (BCDCP) [16] have the good performances in energy and delay metric.

In this work, we propose a two layer hierarchical tree routing protocol called TRPLUCS. TRPLUCS prolongs the lifetime of the system by saving energy, whereas keeps low latency.

The rest of the paper is organized as follows: First we give an overview of the related works in section 2. Then the network and communication models of our proposal are discussed in section 3. The models of tree routing with clustering strategy are described in section 4. A detail description of our protocol TRPLUCS is presented in section 5. Next, in section 6 we present our simulation results compared with other known algorithms. Finally, we make a short conclusion on the work in section 7.

II. RELATED WORKS Various routing protocols have been proposed for

WSNs to alleviate the problems such as low latency and energy efficiency. Among various proposed routing protocols the hierarchical protocols LEACH and PEGASIS provide a good solution to minimize energy consumption and to lengthen network lifetime. LEACH is a cluster-based protocol, while PEGASIS is a chain-based protocol. Our proposed protocol is closely related to two protocols. We strive to leverage the benefits of both protocols while eliminating the drawbacks. Next, we will discuss LEACH and PEGASIS protocol in detail.

A. LEACH Protocol Low-Energy Adaptive Clustering Hierarchy (LEACH)

[13] is a clustering protocol. LEACH reduces the number of nodes communicating directly with the sink. The protocol achieves this by forming different clusters in a self-organizing manner, where each cluster-head collects the data from all the nodes in its cluster, fuses and sends the information to sink. This decision is made by the node n choosing a random number between 0 and 1. If the number is less than a threshold T(n), the node n becomes a cluster-head for the current round.

where P is the desired percentage of cluster-heads, r is the current round, and G is the set of nodes that have not been cluster-heads in the last 1/P rounds. Using this threshold, each node will be a cluster-head at some point within 1/P rounds. LEACH can distribute the energy among the nodes in the network, therefore it enhances system lifetime.

B. PEGASIS Protocol PEGASIS [14] is a chain-based protocol, which makes

further improvement on LEACH. PEGASIS makes a chain routing through greedy algorithm, which is also called the nearest neighbor algorithm. In the chain, each node receives from and transmits to the closest neighbor. The elected chain-leader is responsible for transmitting the final information to the sink.

Greedy chain algorithm begins at a farthest node from the sink, which is the only node in the chain at first. Each terminal node of the chain finds a closest node from the remaining nodes set which are not in the chain. Then the closest node will join the chain and be the new terminal node of the chain. The process repeats till all the nodes join the chain. The greedy chain algorithm in PEGASIS is as follows.

Figure 1. Greedy chain in PEGASIS

Figure 2. Greedy chain algorithm

The main advantages of PEGASIS are: (1) The transmission distances between nodes are minimized. (2) The number of sensor nodes that must send packets to the sink is minimized. The main drawbacks of PEGASIS are: (1) It has excessive delay introduced by the single chain. (2) Greedy algorithm using in PEGASIS is a local search, which cannot provide a global optimal route.

III. NETWORK AND RADIO MODELS

A. Network Model We consider a 300-node network with randomly

distributed nodes in a (200x150) meter area. The sink is located at (x=100, y=300). The length of each signal is 2000 bits and the energy required for data aggregation is Eda=5nJ/bit/signal. Moreover, data processing time per node is taken as 5-10 milliseconds. The radio speed is considered as 2 Mbps.

In the paper, we assume the following properties: (1) Each sensor node has power control and the ability

to transmit data to any other sensor node or directly to the sink.

(2) Our model sensor network contains homogeneous and energy constrained sensor nodes with initial uniform energy.

(3) Every node has location information. (4) There is no mobility.

B. Radio Model In our analysis, we use the same radio model discussed

in [13]. As is the case in [13], we use both the free-space propagation model and the two-ray ground propagation model to approximate the path loss sustained due to wireless channel transmission. Given a threshold transmission distance of d0, the free-space model is employed when d ≤ d0, and the two-ray model is applied for cases where d > d0. Therefore, the energies expended to transmit a k-bit packet to a distance d and to receive that packet with this radio model are as follows.

For transmitter,

ETx(k,d)=Eelec×k+k×Eamp×dn (d>d0, n=4; d≤d0, n=2) (2)

d0 = square (Efs/Emp) (3)

For receiver,

ERx(k) = Eelec×k (4)

Here, Eelec is energy required in transmitter or receiver electronics. Eamp is the energy required by the transmit amplifier to maintain an acceptable SNR. Efs and Emp denote transmit amplifier parameters corresponding to the free-space and the two-ray models, respectively. d0 is the threshold distance of two models.

We make the assumption that the radio channel is symmetric such that the energy required to transmit a message from node A to node B is the same as that from node B to node A for a given SNR. For our environment, we also assume that all the sensors are sensing the environment at a fixed rate and thus always have data to send to the end-user.

IV. TREE ROUTING WITH CLUSTERING STRATEGY MODEL

There are three methods in tree routing protocol in WSNs. The first scheme does not make clustering, while the other two schemes combine clustering strategy with tree routing algorithm. Clustering strategy can reduce low

Procedure ConstructGreedyChain(N,END) 1.Begin 2. N={all nodes}; 3. END = farthest node from SINK; 4. chain= {END}; 5. N=N-{END}; 6. if(N!=NULL) 7. { 8. END=FindCloseNode(N,END); 9. Append(chain,END); 10. goto 5. 11. } 12.END

e0,otherwis)][r mod(1/PP1

PT(n) ∈

⎪⎩

⎪⎨⎧

−= *

latency greatly, and tree routing algorithm can improve energy efficiency.

A. Scheme 1: Tree Construction among All Nodes All the nodes in WSNs construct a tree which may be

the minimum spanning tree or optimal tree. Then data will be transmitted along the tree until it reaches sink. From global view, the energy consumption will be minimal while its delay is much.

B. Scheme 2: Clustering Strategy + Direct Transmission in Cluster + Tree Routing among Cluster-heads

First, all the nodes make clustering. Then tree construction is made among cluster-heads. After that, each node in cluster sends data to its cluster-head directly, while each cluster-head fuses and transmits the data along the tree among the cluster-heads. With the clustering strategy, the second scheme reduces the delay greatly than the first scheme.

C. Scheme 3: Clustering Strategy + Tree Routing in Cluster + Tree Routing among Cluster-heads

First, all the nodes make clustering. Then tree construction is made in each cluster and among all cluster-heads, respectively. After that, each node sends data to its cluster-head along the lowest tree, while each cluster-head fuses and transmits the data along the top tree among the cluster-heads. By means of tree routing instead of direct transmission inside cluster, the third scheme can further improve energy efficiency than the second scheme.

In the following sections, a protocol called TRPLUCS based on the third scheme is proposed in detail. Simulations show it can give a good performance in minimize energy*delay cost.

V. TRPLUCS PROTOCOL In WSNs, energy efficiency and low latency are

considered as two key issues in designing routing protocol. This paper proposes a two layer hierarchical tree routing protocol with location-based uniformly clustering strategy in WSNs (TRPLUCS), which can gives a good compromise between energy consumption and delay. Simulation results show that TRPLUCS performs better than PEGASIS and LEACH in terms of energy*delay cost metric.

The operation of TRPLUCS can be divided by three phases: clustering phase, tree routing construction phase, and data gathering phase. Network routing begins with the formation of clusters. First, TRPLUCS makes balanced clustering based on geographic regions. All clusters are uniformly distributed in WSNs with approximately equal number of nodes in each cluster. Second, TRPLUCS constructs a spanning tree in each cluster and an optimal tree among all the cluster-heads. Third, data is fused and transferred to the cluster-head along the spanning tree in cluster and to the sink along the optimal tree.

In the following subsections we discuss them in details.

Figure 3. TRPLUCS protocol algorithm.

A. Location-based Clustering Strategy In TRPLUCS, the main activities in this phase are

cluster setup and cluster-head election. Location-based clustering algorithm first splits the network area into many equal small squares. These small squares should cover the monitoring area as much as possible. Meanwhile, the side length of small square should meet the conditions as equations below. The number of small squares should be approximate to the desired number of clusters. Then we search the centre of small squares. A node which is closest to the center in each square is selected to be the temporary cluster-head of that cluster. After that, the other node chooses the closest cluster-head to join. At last, a node which has the most residual energy in each cluster is selected to be the current cluster-head. The cluster splitting algorithm ensures that the selected cluster-heads are uniformly placed throughout the whole sensor field. Furthermore, the resulting sub-clusters have approximately the same number of sensor nodes. Accordingly, the cluster setup phase consists of the following five steps:

Step1: Split the network area into equal small squares. Step2: Search the centre of small squares. Step3: Elect temporary cluster-head. Step4: Join cluster according to the shortest distance.

void Location-basedClustering ( ) {

split network area into equal small squares; search centre of small squares; elect temporary cluster-head; join cluster according to shortest distance; elect current cluster-head.

void TreeConstructing ( ) {

construct spanning tree in each cluster; construct optimal tree among cluster-heads;

void TreeRouting ( ) {

spanning tree routing in each cluster; optimal tree routing among cluster-heads; transmit data to sink;

void main ( ) {

for (;;) {

Location-basedClustering ( ); TreeConstructing ( );

TreeRouting ( ); Round++;

Step5: Elect current cluster-head according to the residual energy.

0.9×xm<numx×a<1.1×xm (5)

0.9×ym<numy×a<1.1×ym (6)

0.8×P×N<numx×numy<1.2×P×N (7)

where xm and ym are width and length of WSNs area respectively, numx*numy are the number of small squares, a is the side length of small square, P is the desired percentage of the cluster-heads, and N is the total number of the nodes in WSNs.

Figure 4. Balanced clustering method in TRPLUCS

B. Spanning Tree Construction in Cluster The routing construction in the cluster is computed

using Prim’s minimum spanning tree algorithm where the cluster-head is the root. The algorithm works as follows: Initially, we put a node in the tree which is the cluster-head in our case. After that, in each iteration we select the minimum weighted edge from a vertex in the tree to a vertex not in the tree, and add that edge to the tree. In our case this means that the vertex just included in the tree will send its data through that edge. We repeat this procedure until all the nodes in the cluster are added to the tree. The running time complexity of the algorithm is O(n2) assuming there are n nodes in the cluster. Thus, by computing a minimum spanning tree in the cluster with the energy consumption cost and by routing packets according to that spanning tree, we achieve a minimum energy consuming subsystem in the cluster.

In the process of spanning tree construction, the weight value of two nodes is either:

Costij = 2×k×Eelec+k×Eamp×d2ij (8)

Costis = k×Eelec+k×Eamp×d2is (9)

where Costij is the cost of transmission between node i and j, Costis is the cost of transmission between node i

and sink. dij is the distance between node i and j, and dis is the distance between node i and sink. k is the packet size.

Figure 5. Spanning tree construction algorithm in cluster of TRPLUCS

C. Optimal Tree Construction among Cluster-heads This phase is made up of three steps: Step1: Neighbor Discovery. Each cluster-head makes

its neighbor table. Step2: Cost Computation. Each cluster-head computes

its minimum neighbor cost. Step3: Father Choice. Depending on the cost value,

each cluster-head selects a parent for data transmission. Thus the optimal tree among all the cluster-heads is constructed.

To construct optimal tree, the sink broadcasts information message of cluster-heads include node ID, current residual energy, distance from the sink, distance between cluster-heads. Each cluster-head receives cluster-head information message and makes its neighbor table. A neighbor table contains the neighbor information, such as neighbor node ID, current residual energy, distance from sink, and distance from itself. The computation of father selecting are all based on the information recorded in neighbor table. The neighbor of node i, which has a shorter distance to the sink than node i and has the minimum cost value computed by equations below, will be selected to be as the current father of node i. As described in the equation, the cost of communication between two nodes is related not only with energy consumption of both sides, but also with residual energy of two nodes. If the cluster-head is very close to the sink, then the sink is selected as its father.

Costij =

w×(k×Eelec+k×Eamp×d2ij)/Eni+(1-w)×k×Eelec/Enj (10)

struct { VertexType adjvex;

VRType lowcost; } closedge[MAX_VERTEX_NUM];

Void MiniSpanTree(MGraph G, VertexType u) {

k=LocateVex (G,u ); for (j=0;j<G.vexnum;++j) if(j!=k) closedge[j]={u,G.arcs[k][j].adj}; closedge[k].lowcost=0; for (i=1;i<G.vexnum;++i) { k=minimum(closedge); printf(closedge[k].adjvex,G.vexs[k]); closedge[k].lowcost=0; for(j=0;j<G.vexnum;++j)

if(G.arcs[k][j].adj<closedge[j].lowcost) closedge[j]={G.vexs[k],G.arcs[k][j].adj};

1211109

48.79m

where w is a weighted constant between (0, 1). k is the packet size. Eni is the residual energy of node i. Enj is the residual energy of node j. dij is the distance between node i and j.

After optimal tree construction, each leaf node transmits the data to the selected father. A father node fuses its data with all the data from its children and transmit to its father. At last, the data will be fused and transmitted to the sink, i.e. the root of tree.

TABLE I. NEIGHBOR TABLE

Figure 6. Optimal tree construction algorithm among cluster-heads in TRPLUCS

VI. SIMULATION AND ANALYSIS We make a comparative analysis of clustering structure,

time consumption, energy consumption, system lifetime and energy*delay cost in LEACH, PEGASIS and TRPLUCS protocol. In our experiment, the network and radio models are as mentioned above. Several simulations parameter variations in our test schemes are as follows. The parameters of experiment is the same in three protocol tests.

We run the simulations for about 2000 rounds. At last, we make a conclusion that TRPLUCS shows a better performance in energy*delay metric than PEGASIS and LEACH.

TABLE II. SYSTEM SETTING

A. Clustering Structure Comparison We can see that routing map by PEGASIS shows many

crosses, and many of these crosses are long-range. In LEACH, clusters are made randomly, which make node numbers in each cluster are obviously unequal. Since network delay depends on node numbers in the biggest cluster, the delay will be very much in LEACH. But there is no similar case in routing map by TRPLUCS. In figure, the blue nodes are common nodes. The red nodes are the cluster-heads. The blue color tree in each cluster is the spanning tree constructed by the common nodes. The red color tree is the optimal tree among cluster-heads. We can see, all the clusters are uniformly distributed in WSNs with approximately equal number of nodes in each. Also in TRPLUCS, the transmission distances between nodes are greatly minimized using tree. There are no long-range crosses between two neighbor nodes.

Figure 7. LEACH protocol.

Figure 8. PEGASIS protocol.

Neighbor ID

Residual energy (J)

Distance to sink (m)

Distance to itself (m)

1 0.2 50 58

2 0.1 30.5 38

Parameter Value Initial energy 0.3J

Eelec 50nJ/bit Eamp 100 pJ/bit/m2

Packet length 2000bits Node number 300

Monitoring area 200m×150m Sink location (100m, 300m)

for i=1:1:cluster if CNode (i).state==1

CNode (i).f = 0; CNode (i).d= 0; mincost =inf ; for j=1:1:cluster-1

if CNode(i).d>CNode(j).d&&CNode(i).d > d(i,j)

Cost=w*(k*Eelec + k*Eamp*d2ij)/Eni+

(1-w) *k*Eelec/ Enj; if cost <mincost

mincost =cost; CNode(i).f=j; CNode(i).d=d(i,j);

end if end if

end for end if

end for

Figure 9. Clustering structure in TRPLUCS

Figure 10. Spanning tree construction in cluster of TRPLUCS

Figure 11. Optimal tree construction among cluster-heads in TRPLUCS

B. Time Consumption Comparison We can see that TRPLUCS reduces the data delivery

time greatly than PEGASIS and LEACH. This advantage is achieved by uniformly location-based clustering strategy in TRPLUCS.

Assume that 1 unit time is needed for one node to transmit k bit message with v bps to the neighbor node. Therefore, each unit of delay will meet the equations below.

delay(unit) = k/v (11)

In PEGASIS, there is only one single chain. There need 299 unit time to collect data in the chain and 1 unit time for chain leader to sink. Therefore, in total 300 unit time of delay may occur in one round.

In LEACH, assume n clusters with Si nodes in each cluster (i�[1, n]). In our test we need max(Si-1) unit time in the cluster routing. Then additional n unit time need for cluster-heads routing to sink. So, in total max(Si-1)+n unit time is needed in LEACH. Since the biggest cluster is different in each round, delay value is also changing.

In TRPLUCS, delay computation is the same as that in LEACH. But the nodes number in each cluster is very close due to balanced clustering which makes max(Si-1) is less than that in LEACH. In our test, 44 unit time may require for transmission in one round of TRPLUCS.

Figure 12. Time consumption in TRPLUCS, LEACH, PEGASIS

C. Energy Consumption Comparison We can see that TRPLUCS reduces energy

consumption greatly than LEACH, which can prolong the lifetime of WSNs. This improvement is mainly achieved by two hierarchical tree routing structure in TRPLUCS instead of direct transmission strategy in LEACH. But energy consumption in TRPLUCS is very close to that in PEGASIS. This is due to TRPLUCS gives a good compromise between energy consumption and delay.

Figure 13. Energy consumption in TRPLUCS, LEACH, PEGASIS

D. System Lifetime Comparison Figure below shows the total numbers of nodes that

remain alive over the simulation time. The result shows that TRPLUCS performs better than LEACH. In LEACH, the first node death occurs after 964 rounds. And near to 1200 rounds, almost all the nodes are dead in LEACH. While in TRPLUCS, the first node dies after 1206 rounds. In PEGASIS, there are dead nodes only after 1431 rounds. So PEGASIS performs best.

Figure 14. Number of nodes alive in TRPLUCS, LEACH, PEGASIS E. Energy* Delay Cost Comparison

We can see that TRPLUCS performs best for energy*delay cost. This improvement is mainly achieved by hierarchical tree algorithm in TRPLUCS which can reduce energy consumption and clustering strategy which can reduce time latency greatly.

PEGASIS consumes least energy, while its time latency is very much. LEACH has small delay cost, while energy consumption is much. TRPLUCS gives the best result of energy*delay cost by reaching a tradeoff between energy consumption and time latency.

Figure 15. Energy*Delay cost in TRPLUCS, LEACH, PEGASIS

TABLE III. ENERGY*DELAY COST IN TRPLUCS, LEACH, PEGASIS

VII. CONCLUSION In this paper, we propose TRPLUCS, a clustering

protocols based on two layer hierarchical tree routing scheme. Simulation results show that our protocol outperforms previous approaches, PEGASIS and LEACH in terms of energy*delay cost metric. This is done by constructing spanning tree and optimal tree with balancing cluster technique. Minimizing the total energy of the system while distributing the cluster uniformly has a great impact on system lifetime and performance. Later, we will make use of fast distributed approximation algorithm for spanning tree and more fine clustering method in the routing protocol of WSNs.

ACKNOWLEDGMENT

The project is supported by the Research Foundation of Education Bureau of Hubei Province, China (No. Q20091501).

REFERENCES

[1] A. A. B. Almamou, R. Wrede, P. Kumar, H. Labiodl, and J. Schiller, “Performance evaluation of routing protocols in a real-world WSN”, Proceedings of 2009 Global Information Infrastructure Symposium, GIIS '09, 2009, pp. 1-5.

[2] K. Akkaya and M. Younis, “A Survey of Routing Protocols in Wireless Sensor Networks”, Elsevier Ad-Hoc Network Journal, Vol. 3/3 , 2005, pp. 325-349.

Round Protocol Energy (J)

Delay (unit)

Energy*delay (J×unit)

LEACH 0.0779 52 4.0491

PEGASIS 0.0410 300 12.2963

TRPLUCS 0.0692 44 3.0448

LEACH 0.0796 54 4.2983

PEGASIS 0.0623 300 18.6912

TRPLUCS 0.0579 44 2.5476

LEACH 0.0777 62 4.8200

PEGASIS 0.0512 300 15.3618

TRPLUCS 0.0787 44 3.4628

LEACH 0.0797 53 4.2261

PEGASIS 0.0410 300 12.2963

TRPLUCS 0.0574 44 2.5256

LEACH 0.6794 61 4.8433

PEGASIS 0.0623 300 18.6912

TRPLUCS 0.0661 44 2.9084

[3] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “Wireless sensor networks: A survey”, Computer Networks, vol. 38, no. 4, 2002, pp. 393-422.

[4] R. Szewczyk, J. Polastre, A. Mainwaring, and D. Culler, “Lessons From A Sensor Network Expedition”, 1st European Workshop on Wireless Sensor Networks (EWSN '04), Germany, Jan. 2004, pp. 307-322.

[5] Jamal N. Al-Karaki and Ahmed E. Kamal, “Routing Techniques in Wireless Sensor Networks: A Survey”, IEEE Wireless Communications, vol. 11, no. 6, Dec. 2004, pp. 6-23.

[6] K. Sohrabi, J. Gao, V. Ailawadhi, and G. J. Pottie, “Protocols for self-organization of a wireless sensor network”, IEEE Pers. Commun., vol. 7, no. 5, Oct. 2000, pp. 16–27.

[7] A. Manjeshwar and D. P. Agrawal, “TEEN: a routing protocol for enhanced efficiency in wireless sensor networks”, Proceedings of 15th International Parallel and Distributed Processing Symposium, IPDPS 2001, 2001, pp. 2009-2015.

[8] S. Kim, S. Park, S. Park, and S. Kim, “Energy Efficient Online Routing Algorithm for QoS-Sensitive Sensor Networks”, IEICE Transactions on Communications, vol. E91-B, no. 7, July 2008, pp. 2401-2404.

[9] D. Choi, S. Moh, and I. Chung, “Variable area routing protocol in WSNs : A hybrid, energy-efficient approach”, Proceedings of 10th IEEE International Conference on High Performance Computing and Communications, HPCC 2008, 2008, pp. 397-403.

[10] M. Rajiullah and S. Shimamoto, “An energy-aware periodical data gathering protocol using deterministic clustering in Wireless Sensor Networks (WSN)”, Proceedings of IEEE Wireless Communications and Networking Conference, WCNC 2007, pp. 3016-3020.

[11] Nahdia Tabassum, A K M Ahsanul Haque, and Yoshiyori Urano, “GSEN: An efficient energy consumption routing scheme for wireless sensor network”, ICNICONSMCL'06, pp. 117–122.

[12] W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan, “Energy-efficient Communication Protocols for Wireless Microsensor Networks”, Proceedings of the 33rd Hawaii International Conference on System Sciences, Jan. 2000, pp. 3005-3014.

[13] W. B. Heinzelman, A. Chandrakasan, and H. Balakrishanan, “An Application-Specific Protocol Architechture for Wireless Microsensor Networks”, IEEE Trans. Wireless Commun., vol. 1, no. 4, Oct. 2002, pp. 660-670.

[14] S. Lindsey, C. Raghavendra, and K. M. Sivalingam, “Data Gathering Algorithms in Sensor Networks using Energy Metrics”, IEEE Trans. Parallel and Distrib. Sys., vol. 13, no. 9, Sept. 2002, pp. 924–935.

[15] N. Tabassum, Q. E. K. Mamun and Y. Urano, “COSEN: A Chain Oriented Sensor Network for Efficient Data Collection”, Proceedings of the third International Conference on Information Technology: New Generations, ITNG 2006, 2006, pp. 262-267.

[16] S. D. Muruganathan, C. F. D. Ma, R. I. Bhasin, and A. O. Fapojuwo, “A centralized energy-efficient routing protocol for wireless sensor networks”, IEEE Communications Magazine, vol. 43, no. 3, March 2005, pp. 8-13.

Gengsheng Zheng received the B.S. degree in communication engineering from Xi’an Technology University, China, in 1994 and the M.S. and Ph.D. degrees in computer science from Wuhan University, China, in 1999 and 2006, respectively. Since January 2007, he has been an associate professor in the school of computer science and engineering, wuhan institute of technology, china. His current research interests lie in the area of network and embedded system.

A SVM Method for P2P Traffic Identification based on Multiple Traffic Mode

Hongwei Chen, Xin Zhou, Fangping You, Hui Xu, Chunzhi Wang, Zhiwei Ye

School of Computer Science and Technology, Hubei University of Technology, Wuhan, China Email: chw2001@sina.com

Xin Zhou

Computer Network Center, Hubei University of Technology, Wuhan, China Email: chw75@sohu.com

Abstract—Support Vector Machines (SVM) algorithms are one of the algorithms currently applied in Deep Traffic Inspection (DFI) technologies. This paper realizes online real-time traffic information detection, provides a P2P traffic identification system that supports online SVM analysis and offline SVM training function, and demonstrates the thinking of different identification for IP data traffic and IP-Port data traffic. This paper designs different combinations of traffic features for IP data traffic and IP-Port data traffic, analyzes the effectiveness and exactness of these combinations from various function criteria, and based on a lot of experiments, obtains a best SVM kernel function and a combination of parameters that matches the very combination of traffic features. Index Terms—Peer-to-Peer, Support Vector Machines, Deep Traffic Inspection, Traffic Identification

I. INTRODUCTION

P2P technologies have become one of main technologies that widely used in current network applications, such as communication, entertainment and sharing, due to its advantages that are convenience, high-speed, rich resources and no-center [1]. However, a lot of problems arise at the same time, such as rampant piracy, low quality and leakage of sensitive information [2]. Thus, a lot of studies have done to effectively identify P2P data traffic, which can be divided into two types of technologies that are Deep Packet Inspection (DPI) technologies and the Deep Traffic Inspection (DFI) technologies [3]. The former one tries to judge whether it has found known P2P characteristics through byte-by-byte scan of message contents for data traffics, while the latter one tries to judge whether it has satisfied P2P traffic features through statistical analysis of data traffic [4].

DPI technologies have a high degree of exactness than DFI ones, and they can judge specific application types, but they execute in a low speed, can do nothing to encrypted data and new P2P applications with unknown characteristics, and have a high maintain cost [5].

Compared to DPI technologies, DFI ones execute in a high speed, are effective to encrypted data and new P2P applications with unknown characteristics, and have a relatively low maintain cost, but their exactness is lower than that of DPI ones, and cannot judge specific application types. Currently, studies and products based on DPI technologies are comparably more. However, with the development of anti-identify technologies for P2P software, single use of DPI technologies may not satisfy the requirement. Thus due to their heuristic ability, DFI technologies that based on the identification of traffic features come to the fore, attracting more and more researchers and research centers.

At present, DFI is one of the effective ways to identify P2P traffic, which uses some kinds of mathematical approach to analyze and category features of traffic collected in order to identity P2P traffic [6]. DFI could use many classification algorithms, such as Bayesian decision [7], Neural Net [8], Support Vector Machines [9], and so on. And current studies on the SVM application in the P2P traffic recognition at home and abroad are all offline ones. These researches use software such as ethereal, NetTraffic and Sniffer to obtain the basic information about current network traffic, and handle the information to get the needed one for traffic features, which is finally used for offline training and analysis. In this research process, the system for collecting samples and the one for recognizing P2P traffics are separate, with no direction relationship [10].

According to the classification of study objects, these researches can be divided into three types as follows.

(1) Study on how to configure and select effective traffic features, for the case of known SVM parameter combination [11];

(2) Study on the combination of parameters for SVM algorithm, for the case of a specific traffic information combination;

(3) Study on improving SVM algorithm of a specific parameter, for the case of a specific combination of traffic features.

As for the first kind of studies, there is no unified standardization to configure and select effective traffic features. These studies only discusses the case of a

This work was supported by Natural Science Foundation of HubeiProvince of China (2009CDB100), and Foundation of Wuhan TwilightPlan Project (201050231084).

specific SVM parameter combination, without the solution to problems such as whether the combination is a best one, and whether this best combination will change with the variation of feature combinations for P2P data traffic [12]. As for the second kind of studies, there is no unified paradigm for SVM kernel function selection and specific parameters [13], and the number of selected parameters by these studies is no more than 6, thus the number of combination styles is less [14]. Furthermore, in the former two types, these researches only analyze the overall traffic generated by IP hosts, but don’t analyze each specific IP-Port traffic according to the practical needs for control [15].

This paper realizes a P2P traffic identification system that unifies online sampling and online SVM analysis for the problem of separation of the sampling system and the analyzing system [16]. The proposed system can realize online real-time statistics of basic traffic information and directly process it to obtain the needed information for traffic features, and this system also support online SVM analysis function and offline SVM training function. We argue that, different combinations of traffic features may lead to different best combinations of SVM parameters, which may influence the effectiveness of the same combination of traffic features, Thus, this paper combines the first type and the second type of studies, discusses the best SVM parameter combinations for different combination of traffic features, and compares the experiment results, analyzing the differences of traffic feature combinations. Additionally, according to the control objects, this paper divides current control technologies for P2P traffic into two types that are bandwidth control and application control. And based on this classification, this paper proposes to divide P2P traffic into IP traffic mode and IP-Port traffic mode for separate identification, studies corresponding combinations of traffic features and best SVM parameters. The whole research relies on the implemented system, tests 10 combination of SVM parameters (the number is 5 to 10) [17], and comparatively studies the performance criteria for different combination styles in cases of different traffic modes and different signatures [18].

II. DFI-ONLINE MONITORING MODEL Fig. 1 is based on the DFI-SVM model of LIBSVM,

including Statistic Module, Distill information Module, On-line Module, Off-line Module, Dynamic Traffic Information statistical structure sets, and Static Traffic Information used to store files, which are introduced as follows.

Statistic Module: The module contains a Traffic Statistics and a Dynamic Traffic Information Sub-module which is responsible for statistics on Net Traffic information and then will be written in the Dynamic Traffic Information Structure sets. (1) Traffic Statistics Sub-module: It’s responsible for statistics on Net Traffic information. (2) Dynamic Traffic Information Sub-module: Dynamic Traffic Information is a kind of data structure which recorded all detected IP host addresses, traffic statistics generated by the host and traffic statistics.

These data are stored in the memory space to support online extracting and offline storage.

Figure 1. DFI-SVM model

Information Extracting Module: According to the users’ choice, IP or IP-Port information would be extracted from the Dynamic Traffic Information and Static Traffic Information, which would be handed to the off-line training and online classification module. The module contains the IP information, IP-Port information and Information Extracting Sub-module. (1) IP Information Sub-module: IP Information is a kind of data structure sets. The original data, which would be obtained from the Dynamic Traffic Information or Static Traffic Information, is disassembled to get IP information, that is, only contains the traffic statistical information generated by the host IP. (2) IP-Port Information Sub-module: IP-Port Information is a kind of data structure sets. The original basic data, which would be obtained from the Dynamic Traffic Information or Static Traffic Information, is disassembled to get IP-Port information, which means that, it only contains the Traffic Data Statistical Information with IP-Port identification. (3) Information Extracting Sub-module: According to the users’ choice, the module is responsible for extracting the specific the type of the original data. After processing data, the corresponding DFI feature data will be obtained and reflected into memory space as the form of documents for off-line training in the use.

Off-line Module: The module contains the Static Traffic Information, Training sets, Default Parameters, Off-line Training and Off-line Classification Sub-module. (1) Static Traffic Information Sub-module: This information save a copy of the Dynamic Traffic Information automatically after the system stopping testing work as the form of Access Database stored. (2) Training Sets Sub-module: The collection is generated by the LIBSVM Training Function as the form of documents stored and would be called in the implementation of Off-line or on-line classification. (3)Default Parameter Sub-module: The sub-module receives user’s specified parameters, to determine the specific configuration of LIBSVM Training Function. (4)Off-line Training Sub-module: The sub-module training function is the training function of the original LIBSVM source, a function of

which loads the sample files that are converted into pointing to memory-mapped file address. (5) Off-line Classification Sub-module: The sub-module is loading the training sets and categories, sample sets in-memory. It’s mainly used to test the training sets and classification accuracy.

On-line Module: The module classifies the data which is submitted by the received Information Extracting Module using training sets. Then when geting the results, it will be recorded to dynamic traffic Information. The module contains the On-line Classify and Classify Sets Sub-module. (1) On-line Classify Sub-module: The sub-module classifies the captured traffic statistics information currently and runs automatically every the specified time. Specifically categorized information extraction is controlled by both Information Extracting Module and Training Sets. (2) Classify Sets Sub-module: The module is responsible for the classified data structuring sets in the memory to write the corresponding dynamic traffic information statistical structuring sets in order to facilitate testing and research.

III. P2P TRAFFIC FEATURES It is found that, current P2P traffic control as the

control object can be divided into two categories: bandwidth control and application control. (1) Bandwidth Control: When some IP host computers generate P2P traffic, it would restrict the net bandwidth that the IP owned. That means all network applications that IP has are subject to receive the same network bandwidth restrictions as punishment. (2) Application Control: When they find P2P traffic exists in the IP-Port traffic maintained by some IP host computer, then IP-Port traffic will be restricted to be transmitted.

For the two kinds of applications above, we proposes the distinctive recognition between IP traffic features and IP-Port traffic features, in order to provide more scientific and accurate identification of P2P services for the two types of control.

IP traffic is the set of all data streams generated by IP host, including large P2P traffic and Non-P2P traffic. Non-P2P traffic may make detection difficult, which having a negative impact on P2P traffic features. So it is necessary to set multiple traffic features to improve the recognition accuracy. IP traffic comprises IP traffic, IP-Port traffic, and IP traffic features are more obvious than that of IP-Port when the overall traffic is relatively small.

IP-Port traffic is a stream of data generated by a specific port. Generally, when an application opens a port that will be occupied until the application is released. Compared with IP traffic, there is no other interfere for data stream with IP-Port traffic, and IP-Port traffic is very pure. When the overall traffic is relatively small, IP-Port traffic is very small, and it needs to use mathematical methods for processing effective detection.

Table I is IP traffic features, and Table II is IP-Port traffic features.

TABLE I. IP TRAFFIC FEATURES

Traffic Feature Explanation

ARPn Number of ARP Request Packets

TCPIO Ratio of Intraffic TCP Traffic to Outtraffic TCP Traffic

UDPIO Ratio of Intraffic UDP Traffic to Outtraffic UDP Traffic

TATIO Ratio of Total Intraffic Traffic to Outtraffic Traffic

AVP Average Traffic Speed AVL Average Packet Length

TUL Ratio of Average TCP Packet Length to that of UDP

TUF Ratio of TCP Traffic to UDP Traffic TCPC Ratio of TCP Traffic to Total Traffic UDPC Ratio of UDP Traffic to Total Traffic

TABLE II. IP-PORT TRAFFIC FEATURES

Traffic Feature Explanation

pT Traffic Duration pAS Average Traffic Speed pAL Average Packet Length

pTUAL Ratio of Average TCP Packet Length to that of UDP

pTUF Ratio of TCP Traffic to UDP Traffic

The reasons of IP traffic features abandoning the duration feature: For an IP host, duration equals test end time subtracts discovery time. When one IP host uses P2P software within a while time, and stops using the network after a very long time, then this feature becomes ineffective.

The reasons of IP traffic features abandoning the number of ARP packets: IP-Port traffic does not use ARP protocol.

The reasons of IP traffic features abandoning the number ratio of send and receive TCP packets, the number ratio of send and receive UDP packets, the proportion of TCP in the traffic and the proportion of UDP in the traffic: IP-Port traffic is single communicating. The same IP-Port traffic cannot simultaneously upload and download.

It is worthy of special mention that, IP-Port traffic features are sets that are composed of several feature subsets such as {pT0, pAS0, pAL0, pTUAL0, pTUF0},{pT1, pAS1, pAL1, pTUAL1, pTUF1}…{ pTn, pASn, pALn, pTUALn, pTUFn}. These subsets make up a matrix when sorting by rows.

pT0 pAS0 pAL0 pTUAL0 pTUF0 pT1 pAS1 pAL1 pTUAL1 pTUF1 …………………………

pTn-1 pASn-1 pALn-1 pTUALn-1 pTUFn-1

pTn pASn pALn pTUALn pTUFn

Compared with IP traffic, IP-Port traffic is relatively small; it is difficult to achieve satisfactory recognition results when only using a single set of data for detection. Therefore, we can use column as unit of feature sets and respectively calculate variance or mean absolute

deviation of data obtain more precise characteristics of value, and make traffic features become evident.

The variance formula is: 2

1( ) ( )1

V x S x xn =

= = −− ∑

(1) )(xV represents variance of x , xS represents standard

deviation of x , n represents the number, and x represents mean of x .

The mean absolute deviation formula is:

AAD x x xn =

= −∑ (2)

x represents the value of feature items, n represents the number, and x represents mean of x .

The calculating result between variance and mean absolute deviation is 0.1~0.3 through practical test. Under the same parameter configuration, the recognition rate of variance formula is higher than that of mean absolute deviation formula. So we use variance formula in this paper.

IV. RBF KERNEL FUNCTION Through tests, we found that whatever kind of SVM

classification, RBF (Radial Basis Function) kernel function can get the highest scores in this study. RBF kernel function is mainly used to solve non-linear separable problems. It fully shows that, traffic statistics sample sets obtained from the net are a separable collection of non-linear. RBF kernel function formula:

( ) ( )2, expK x y x yγ= − − (3)

There into1

So for any given sample sets X={x1, x2,…xi}, i = 1, 2, 3…N, Y={-1, 1}, using the RBF kernel function can be reflected to high-dimensional space, which would be sought optimal classifier face, and using the decision function.

( ) ( ){ } ( ){ }* *sgn sgn ,

nf x w x b yK x y bi ii

α∑= + = +=

The *iα is one positive component of the best solution

in optimal classifier face problem.

( ) ( )1min ,

1 , 12

n nF y y K x yi i j i j i ji i j

α α α αα

∑ ∑= −= =

*ib is the positive component threshold of *

( )* *,

nb y y K x yi i i ii

α∑= −=

It can be seen from (5) and (6), to determine the impact factor of the final classification accuracy rate is C and σ, and C is the punish modulus.

V. TEST RESULTS AND ANALYSIS This paper uses nearly 600 samples tested, which

contains more than 300 IP-Port samples and more than 200 IP samples, and P2P sample is positive class while non-P2P sample is negative one. Meanwhile, using the 10 LIBSVM parameters combinations has been tested in the IP mode, IP-Port Mode and IP & IP-Port mode respectively according to the exhaustive algorithm. And then the corresponding optimal LIBSVM order combinations can be obtained, with the help of which to verify the validity of the traffic feature selection. This test related indicators as followed: Rate (Identification Rate) = Correct Identify Records /

All Records TPR (True Positive Rate) = TP / (TP + FN) TNR (True Negative Rate) = TN / (TN + FP) FPR(False Positive Rate) = FP / (FP + TN) FNR (False Negative Rate) = FN / (TP + FN)

Parameter selection and value range of LIBSVM is described in the following table.

TABLE III. PARAMETER TABLE OF LIBSVM

Parameter Value Range Explanation

{0,1,2,3,4} represents C-SVC, nu-SVC, one-class SVM, e- SVR, and n – SVR

Setting type of SVM

-t {0,1,2,3}represents linear, polynomial, RBF, and sigmoid kernel

Setting type of kernel function

-d {1,2,3,4,5} Setting degree in kernel function

-g {0,1,2,3,4} Setting gamma in kernel function

-r {0,1,2,3,4} Setting coef in kernel function

-c {1,2,3,4,5}

Setting the parameter C of C-SVC, e-SVR, and nu-SVR

-n {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9}

set the parameter nu of nu-SVC, one-class SVM, and nu-SVR

-p {0.1, 0.2, 0.3, 0.4, 0.5} Setting the epsilon in loss function of e-SVR

-e {0.001, 0.005, 0.01, 0.015, 0.02}

Setting tolerance of termination criterion

-wi {1,2,3,4,5}

Setting the parameter C of class i to weight*C, for C-SVC

A. IP mode test

As is shown in Table I, there are 10 traffic feature items in IP mode. As is shown in Fig. 2, each feature item independently judge P2P traffic application in IP mode, the curved shape of recognition accuracy rate is changed with SVM parameter. The SVM parameter is same in the same vertical line. We can conclude that, different feature items have different optimal SVM parameter

combinations, and different SVM parameter combinations also affect the validity of traffic feature items.

Figure 2. Recognition rate of single feature in IP mode

Furthermore, we carry out incremental test of traffic feature combination in IP mode, namely first test traffic feature ARPn, second test traffic features ARPn and TCPIO and so on, until all traffic features are tested. The test result is shown in Table IV.

TABLE IV. INCREMENTAL TEST TABLE OF TRAFFIC FEATURE COMBINATION IN IP MODE

Attribute Number

Rate (%)

TPR (%)

TNR (%)

FPR (%)

FNR (%)

1 82 81.82 0.0 100 50 2 89 89.77 16.67 83.33 50 3 89 89.77 16.67 83.33 50 4 89 89.77 16.67 83.33 50 5 90 90.80 15.38 84.62 50 6 91 91.86 14.29 85.71 50 7 91 91.86 14.29 85.71 50 8 92 92.94 13.33 86.67 50 9 92 92.94 13.33 86.67 50 10 93 94.05 12.5 87.5 50

Table IV provides the incremental traffic feature combinations in accordance with traffic feature sequence in Table I. As shown in Table IV, with the number of feature items increased, the recognition accuracy rate of P2P traffic is increased.

Fig. 3 is recognition rate of combined features in IP mode, which provides the linear formula of recognition rate that is changed with the number of feature items increased. The linear formula is:

Y = 0.8155x +85.418

Figure 3. Recognition rate of combined features in IP mode

Table V provides several feature combinations which have equal recognition rate. In order to further study how the equal feature items affect the overall recognition rate, recognition rate of different eliminated feature combinations in IP mode is provide in Table IV. The test result is shown in Table V:

TABLE V. RECOGNITION RATE OF DIFFERENT ELIMINATED FEATURES COMBINATION IN IP MODE

Eliminated Features

Rate (%)

TPR (%)

TNR (%)

FPR (%)

FNR (%)

TCPIO 90 90.80 15.38 84.62 50 UDPIO 92 92.94 13.33 86.67 50 TATIO 93 94.05 12.5 87.5 50 AVL 88 88.76 18.18 81.82 50 TUL 92 92.94 13.33 86.67 50 TUF 91 91.86 14.29 85.71 50

TCPC 91 91.86 14.29 85.71 50

Figure 4. Recognition rate of different eliminated features combination in IP mode

Table V shows traffic feature Average Packet Length is the greatest impact on the overall recognition accuracy, while the minimal factors is traffic feature Ratio of Total

in-out traffic, because P2P traffic from the test sample data mainly come from P2P file share application such as Thunder and BitTorrent which generate large TCP traffic, and non-P2P data stream interference from UDP protocol.

To sum up, as for the IP mode, the global in-out rate of P2P traffic for identification exactness degree, TPR, FPR and TNR is higher than that of UDP traffic, which indicates that, the global in-out rate for recognizing P2P traffic is a litter better than that of UDP traffic, while as for the case of recognizing non-P2P traffic, it is less exact than that of UDP traffic. In the global test, due to the influence of non-P2P flow, the global flow feature rate plays no important influence on the recognition rate of feature combination for flow. And FPR has an obvious advantage, which is most sensitive to non-P2P flow recognition. This mode uses the C_SVC classification method, degree is 4, g is 1/4, selects the RFB kernel function, and sets 3 for the coef parameter, 3 for penalty factor C, 0.001 for tolerable deviation, 1 for the weight of C. Thus then, the best performance can be obtain, which is that Rate: 93%, TPR: 94.05%, TNR: 12.5%, FPR: 87.5%, FNR: 50%.

B. IP-Port mode test

As is shown in Table III, there are 5 feature items for traffics for the IP-Port mode. First of all, this part does separate tests for these 5 items, with results shown in Figure 5.

Figure 5. Recognition rate of single feature in IP-Port mode

As shown in Fig. 5, each feature item independently judge P2P traffic application in IP-Port mode, the curved shape of recognition accuracy rate is changed with SVM parameter. The SVM parameter is same in the same vertical line. Table VI shows the incremental test of traffic feature combination in IP-Port mode, which follows the combinations of feature items in an ascending order shown in Table II.

TABLE VI. INCREMENTAL TEST TABLE OF TRAFFIC FEATURE COMBINATION IN IP-PORT MODE

Attribute Number

Rate (%)

TPR (%)

TNR (%)

FPR (%)

FNR (%)

1 83.8 86 16.87 83.13 50 2 86.11 95.56 16.37 83.63 50 3 99.54 100 0.68 99.32 0 4 99.54 100 0.68 99.32 0 5 99.54 100 0.68 99.32 0

As is indicated in Table VI, the exactness degree of P2P traffic raises with the increase of feature items, the rule of which is presented by the following linear formula.

Y = 4.3386x +81.706 Figure 6 shows the recognition rate of combined

features in IP-Port mode. Figure 6. Recognition rate of combined features in IP-Port mode

After analyzing Table VI, it can be found that, with the addition of pAL, the recognition rate will rise to 99.51%, and this rate will remain even with addition of other feature items. In order to analyze the influence of these 3 feature items on the global recognition rate for the combination of features. This paper does the exclusion test, with results as follows.

TABLE VII. RECOGNITION RATE OF DIFFERENT ELIMINATED FEATURES COMBINATION IN IP-PORT MODE

Eliminated Features

Rate (%)

TPR (%)

TNR (%)

FPR (%)

FNR (%)

pAL 99.54 100 0.68 99.32 0 pTUAL 99.54 100 0.68 99.32 0 pTUF 99.54 100 0.68 99.32 0 pAL+

pTUAL 99.54 100 0.68 99.32 0

pAL+ pTUF

99.54 100 0.68 99.32 0

pTUAL+ pTUF

99.54 100 0.68 99.32 0

Figure 7. Recognition rate of different eliminated features combination in IP-Port mode

Table VII provides the recognition rate of different eliminated features combination in IP-Port mode. As is shown in Table VII, the combinations of any one of the

three feature items pAL, pTUAL and pTUF, which have equal influences on the global recognition rate, pT and pAS, can all obtain the highest identification exactness degree. We argue that, under the same recognition rate, the use of most effective feature items can satisfy the more exact recognition rate in other experiment conditions.

To sum up, in a single experiment, the feature item for average speeds is the lowest in the identification rate, but this feature has an efficiency of 100% for P2P traffic identification, which is one of the important marks. Among all these feature items, the TCP/UDP traffic rate is the most effective, which is 100% for TPR, 6.45% for TNR, 93.55% for FPR and 0% for FNR. The IP-Port mode uses the C_SVC classification method, selects the RFB kernel function, and sets 3 for the coef parameter, 3 for penalty factor C, 0.005 for tolerable deviation, 1 for the weight of C. Thus then, the best performance can be obtain, which is that rate is 99.54%, TNR is 0.68%, FPR is 99.32% and FNR is 0%.

C. Mode Comparison

Figure 8 shows the performance comparison between IP and IP-Port mode. As is shown in Figure 8, under the experimental conditions presented in this paper, the use of IP-Port mode is obviously better than that of IP-mode in the exactness degree of identification and criteria that are TPR, TNR, FPR and FNR. Especially for the identification of non-P2P data traffics, IP Port is influenced by many interference criteria, with an error rate 50%, while that of IP-Port mode is 0, which indicates that the use of IP-Port mode for the identification of data traffics can effectively avoid the influences of interference criteria.

Figure 8. Performance comparison chart between IP and IP-Port mode

VI. CONCLUSION This paper realizes an online P2P identification system

with the help of LIBSVM, proposes the separation of mining for the IP and IP-Port data traffics, and provides corresponding traffic features. By a large number of actual data testing, we have obtained the combination of SVM commands that can be applied to identification of IP and IP-Port data traffics, and have also verified the effectiveness of the traffic features. As for the test presented in this paper, the exactness degree of the IP-Port

mode is 99.54%, 6.54% higher that of the IP mode. This result indicates that, when considering the identification issue of data traffics, the IP-Port mode is obviously more effective than the IP mode.

However, because of the limitations of the experimental environment, the testing data used in this process still cannot cover all the factors that can affect test results and all the combinations of feature items. On the other hand, issues such as whether there are more effective features and whether existing features presented in this paper would become more effective by some mathematical treatments, calling for further studies.

ACKNOWLEDGMENT We would like to thank the reviewers and editors for

their detailed and valuable comments.

REFERENCES [1] S. Saroiu, P. K. Gummadi, S. D. Gribble, “A

Measurement Study of Peer-to-Peer File Sharing Systems”, Proceeding of Multimedia Computing and Networking, 2002.

[2] D. S. Wallach, “A Survey of Peer-to-Peer Security Issues”, Proceedings of International Symposium on Software Security”, 2002, pp. 42-57.

[3] C. Fivos, M. Panayiotis, “Identifying known and unknown peer-to-peer traffic”, Proceeding of Fifth IEEE International Symposium on Network Computing and Applications, 2006, pp. 93-102.

[4] S. Ohzahata, Y. Hagiwara, M. Terada, K. Kawashima, “A Traffic Identification Method and Evaluations for a Pure P2P Application”, Lecture Notes in Computer Science, Vol. 3431, 2005, pp. 55-68.

[5] C. L. Hayes, Y. Luo, “DPICO: a high speed deep packet inspection engine using compact finite automata”, Proceedings of the 3rd ACM/IEEE Symposium on Architecture for networking and communications systems, 2007.

[6] A. Madhukar, C. Willamson. A longitudinal study of P2P traffic classification , “ Proceeding of 14th IEEE International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunication Systems ”, 2006, pp.179-188.

[7] K. C. Khor, C. Y. Ting. Amnuaisuk, S. P., “Comparing Single and Multiple Bayesian Classifiers Approaches for Network Intrusion Detection”,Proceedings of 2010 Second International Conference on Computer Engineering and Applications, Vol 2, 2010, pp.325-329.

[8] P. Kraipeerapun, “Neural network classification based on quantificaiton of uncertainty”,Murdoch University, 2008.

[9] R. Wang, Y. Liu, Y. Yang et al., “ Solving the app-level classification problem of P2P traffic via optimized support vector machines”, Proceedings of Sixth International Conference on Intelligent

Systems Design and Applications,Vol 2, 2006, pp. 534-539.

[10] F. Constantinou, P. Mavrommatis, “Identifying Known and Unknown Peer-to-Peer Traffic”, Proceedings of the Fifth IEEE International Symposium on Network Computing and Applications, 2006, pp. 93-102.

[11] A. Yang, S. Jiang, H. Deng, “A P2P Network Traffic Classification Method Using SVM”, Proceeding of 9th International Conference for Young Computer Scientists, 2008.

[12] A. McGregor, M. Hall, P. Lorier, J. Brunskill, “Traffic Clustering Using Machine Learning Techniques”, Proceeding of Passive & Active Measurement Workshop 2004, 2004.

[13] M. Adankon, M. Cheriet, “New Formulation of SVM for Model Selection. Neural Networks”, Proceeding of International Joint Conference on Neural Networks, 2006, pp. 1900-1907.

[14] J. Early, C. Brodley, C.Rosenberg, “Behavioral authentication of server traffics”, Proceeding of the 19th Annual Computer Security Applications Conference, 2003.

[15] A. W. Moore, D. Zuev, “Discriminators for use in traffic-based classification”, Technical report, Intel Research, Cambridge, 2005.

[16] Marcell Perényi, et.al., “Identification and Analysis of Peer-to-Peer Traffic”, Journal of Communications, Vol. 1, No. 7, 2006, pp.36-46.

[17] O Chapelle, V Vapnik et al., “Choosing Multiple Parameters for Support Vector Machines”, Machine Learning, Vol. 46, 2002, pp.131-159.

[18] J. Shawetaylor, N. Cristianini, “ Kernel methods for pattern analysis”,Cambridge: Cambridge University Press, 2004.

Hongwei Chen (1975-), male, from Hubei Province, PHD, Associate Professor of Hubei University of Technology, interested in Peer-to-Peer, Grid Computing, Information Security, Mobile Agent.

Xin Zhou(1983-), master of Hubei University of Technology, interested in Peer-to-Peer. Currently working for Network Center in Hubei University of Technology, mainly engaged in network security management and research.

Fangping You(1982-), master of Hubei University of Technology, interested in Peer-to-Peer. Currently working for the laboratory in the School of Electrical and Electronic Engineering, engaged in research and teaching.

Hui Xu (1983-), PHD, Lecturer of Hubei University of Technology, interested in network and service management. Since 2006, she has been a certified computer system analyst in P.R. China. In July 2008, her biography was selected for inclusion in the 26th edition (2009) of the Marquis Who’s Who in the World, California, USA.

Chunzhi Wang(1963-), female, from Hubei province of China, Master's degree, Professor of Hubei University of Technology, Dean of School of Computer Science, interested in the security of network and computer network, Computer supported cooperative work. The Chairman of Wuhan of CCF Young Computer Scientists & Engineers Forum(2010).

Zhiwei Ye (1978-), male, from Hubei Province, PHD, Associate Professor of Hubei University of Technology, interested in Computational Intelligence, Image Processing.

Call for Papers and Special Issues

Aims and Scope. Journal of Networks (JNW, ISSN 1796-2056) is a scholarly peer-reviewed international scientific journal published monthly, focusing on theories,

methods, and applications in networks. It provide a high profile, leading edge forum for academic researchers, industrial professionals, engineers, consultants, managers, educators and policy makers working in the field to contribute and disseminate innovative new work on networks.

The Journal of Networks reflects the multidisciplinary nature of communications networks. It is committed to the timely publication of high-

quality papers that advance the state-of-the-art and practical applications of communication networks. Both theoretical research contributions (presenting new techniques, concepts, or analyses) and applied contributions (reporting on experiences and experiments with actual systems) and tutorial expositions of permanent reference value are published. The topics covered by this journal include, but not limited to, the following topics:

• Network Technologies, Services and Applications, Network Operations and Management, Network Architecture and Design • Next Generation Networks, Next Generation Mobile Networks • Communication Protocols and Theory, Signal Processing for Communications, Formal Methods in Communication Protocols • Multimedia Communications, Communications QoS • Information, Communications and Network Security, Reliability and Performance Modeling • Network Access, Error Recovery, Routing, Congestion, and Flow Control • BAN, PAN, LAN, MAN, WAN, Internet, Network Interconnections, Broadband and Very High Rate Networks, • Wireless Communications & Networking, Bluetooth, IrDA, RFID, WLAN, WMAX, 3G, Wireless Ad Hoc and Sensor Networks • Data Networks and Telephone Networks, Optical Systems and Networks, Satellite and Space Communications

Special Issue Guidelines Special issues feature specifically aimed and targeted topics of interest contributed by authors responding to a particular Call for Papers or by

invitation, edited by guest editor(s). We encourage you to submit proposals for creating special issues in areas that are of interest to the Journal. Preference will be given to proposals that cover some unique aspect of the technology and ones that include subjects that are timely and useful to the readers of the Journal. A Special Issue is typically made of 10 to 15 papers, with each paper 8 to 12 pages of length.

The following information should be included as part of the proposal: • Proposed title for the Special Issue • Description of the topic area to be focused upon and justification • Review process for the selection and rejection of papers. • Name, contact, position, affiliation, and biography of the Guest Editor(s) • List of potential reviewers • Potential authors to the issue • Tentative time-table for the call for papers and reviews If a proposal is accepted, the guest editor will be responsible for: • Preparing the “Call for Papers” to be included on the Journal’s Web site. • Distribution of the Call for Papers broadly to various mailing lists and sites. • Getting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. Authors should be

informed the Instructions for Authors. • Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact

information. • Writing a one- or two-page introductory editorial to be published in the Special Issue.

Special Issue for a Conference/Workshop A special issue for a Conference/Workshop is usually released in association with the committee members of the Conference/Workshop like

general chairs and/or program chairs who are appointed as the Guest Editors of the Special Issue. Special Issue for a Conference/Workshop is typically made of 10 to 15 papers, with each paper 8 to 12 pages of length.

Guest Editors are involved in the following steps in guest-editing a Special Issue based on a Conference/Workshop: • Selecting a Title for the Special Issue, e.g. “Special Issue: Selected Best Papers of XYZ Conference”. • Sending us a formal “Letter of Intent” for the Special Issue. • Creating a “Call for Papers” for the Special Issue, posting it on the conference web site, and publicizing it to the conference attendees.

Information about the Journal and Academy Publisher can be included in the Call for Papers. • Establishing criteria for paper selection/rejections. The papers can be nominated based on multiple criteria, e.g. rank in review process plus

the evaluation from the Session Chairs and the feedback from the Conference attendees. • Selecting and inviting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors.

Authors should be informed the Author Instructions. Usually, the Proceedings manuscripts should be expanded and enhanced. • Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact

information. • Writing a one- or two-page introductory editorial to be published in the Special Issue. More information is available on the web site at http://www.academypublisher.com/jnw/.

(Contents Continued from Back Cover)

Reducing Complexity and Consumption in Future Networks G. M. Tosi Beleffi, G. Incerti, L. Porcari, S. Di Bartolo, M. Guglielmucci, A. L. J. Teixeira, L. Costa, N. Wada, J. Prat, J. Lazaro, and P. Chanclou Decrease of the Link PMD by Fiber Exchange and Investigation of the PMD Distribution along Buried Optical Fibers with a POTDR Armin Ehrhardt, Manuel Paul, Lars Schürer, Christoph Gerlach, Wolfgang Krönert, Daniel Fritzsche, Dirk Breuer, Volker Fürst, Normand Cyr, Hongxin Chen, Gregory W. Schinn Deployment and Validation of GMPLS-Controlled Multi-layer Integrated Routing over the ASON/GMPLS CARISMA Test-bed Fernando Agraz, Luis Velasco, Jordi Perelló, Marc Ruiz, Salvatore Spadaro, Gabriel Junyent, and Jaume Comellas Enhancing Performance of Optical Communication Systems with Advanced Optical Signal Processing Ivan Glesk, Marc Sorel, Anthony E. Kelly, and Paul R. Prucnal Advanced Test-beds to Validate Physical Estimators in Heterogeneous Long Haul Transparent Optical Networks Annalisa Morea, Florence Leplingard, Jean-Christophe Antona, Pascal Henri, Thierry Zami, and Daniel C. Kilper All-optical Label Swapping Techniques for Optical Packets at Bit-rate Beyond 160 Gb/s Nicola Calabretta, Hyun-Do Jung, and Harm Dorren Tb/s Transmission and Routing Systems Using Integrated Micro-Photonic Components Efstratios Kehayas, Leontios Stampoulidis, and Paraskevas Bakopoulos

REGULAR PAPERS A Peer-to-Peer Game Model using Punishment Strategies Chunzhi Wang, Hongwei Chen, Ke Zhou, Hui Xu, and Zhiwei Ye Reliable Resource Search in Scale Free Peer-to-Peer Network Wei Song, Wenbin Hu, Zhengbing Hu, and Xi Zeng Tree Routing Protocol with Location-based Uniformly Clustering Strategy in WSNs Gengsheng Zheng and Zhengbing Hu A SVM Method for P2P Traffic Identification based on Multiple Traffic Mode Hongwei Chen, Xin Zhou, Fangping You, Hui Xu, Chunzhi Wang, and Zhiwei Ye

Journal of Networks - CiteSeerX

Documents

Transcript of Journal of Networks - CiteSeerX

Journal of Biomedical Informatics - CiteSeerX

core networks, social isolation, and new media - CiteSeerX

Advances in Computer Science: an International Journal - CiteSeerX

IP Flow Mobility in PMIPv6 Based Networks - CiteSeerX

Journal of Communications Contents - CiteSeerX

Information Theory of Complex Networks - CiteSeerX

Mediterranean Journal of Social Sciences - CiteSeerX

European Journal of Scientific Research - CiteSeerX

and Materials Journal of Sandwich Structures - CiteSeerX

Internationally indexed journal Internationally ... - CiteSeerX

Pan Arab Journal of Oncology - CiteSeerX

Algorithms for analysis of large networks - CiteSeerX

Journal of Computers - CiteSeerX

Attacks on Peer-to-Peer Networks - CiteSeerX

The Journal of Alternative Education, 1998. 550p. - CiteSeerX

Security Analysis of Social Networks - CiteSeerX

CMRD JOURNAL OF MANAGEMENT RESEARCH - CiteSeerX

EUROPEAN JOURNAL OF SOCIAL SCIENCES ISSN ... - CiteSeerX

The Iranian EFL Journal - CiteSeerX

veterinary - journal - CiteSeerX