Photon Netw CommunDOI 10.1007/s11107-009-0193-y

Layer-preference policies in multi-layer GMPLS networks

Péter Fodor · Gábor Enyedi · Gábor Rétvári ·Tibor Cinkler

Received: 5 November 2008 / Accepted: 3 February 2009© Springer Science+Business Media, LLC 2009

Abstract We address the problem of routing Label Swit-ched Paths (LSPs) in multi-layer networks based on the Gen-eralized MultiProtocol Label Switching (GMPLS) paradigm.In particular, we pursue policies for choosing the appropri-ate layer to host a new LSP request, as we find that suchlayer-preference policies have significant impact on networkperformance. We discuss several simple layer-preference pol-icies and we reveal why these simple policies ruin networkperformance in the long run. Consequently, we develop anefficient heuristics, the Min-phys-hop routing and wavelengthassignment algorithm, to govern the selection of the best layerof a multi-layer network in which to host new LSP requests.We discuss the applicability of this algorithm with respect tothe state-of-the-art GMPLS standards, above all, the GMPLSrouting extensions to OSPF-TE. By extensive simulations,we justify that the Min-phys-hop algorithm produces close-to-optimal blocking and resource consumption under almostall possible selections of input parameters, and this is regard-less of the wavelength and Optical-Electrical-Optical (OEO)conversion capability present in the network.

Keywords Multi-Layer Traffic Engineering · GMPLS

P. Fodor (B) · G. Enyedi · G. Rétvári · T. CinklerDepartment of Telecommunications and Media Informatics, BudapestUniversity of Technology and Economics, 2 Magyar tudósok körútja,H-1117 Budapest, Hungarye-mail: fodorp@tmit.bme.hu

G. Enyedie-mail: enyedi@tmit.bme.hu

G. Rétvárie-mail: retvari@tmit.bme.hu

T. Cinklere-mail: cinkler@tmit.bme.hu

1 Introduction

The Term Generalized MPLS (GMPLS) signifies the archi-tecture, in which a number of switched network transportlayers are stacked onto each other and are operated under theauthority of a unified control function.

Traditionally, different technological layers of multi-layernetworks were operated by isolated control planes with no,or very limited information exchange between the controlplanes responsible for the different layers. This model iscalled the Overlay model, since the upper layer is simplyoverlayed on top of the lower layer without the two beingaware of each other in any regards. This model was laterextended to allow for limited information exchange betweencontrol planes. The resultant control architecture is called theAugmented model. With the advent of GMPLS, it becamepossible to completely separate the control plane from thedata plane, which opened the way to introduce all the tech-nological layers to under the authority of a unified controlplane. This huge integration of vastly different network tech-nologies is made possible by the abstraction of the notion of“labels”: basically any quantity of traffic flow that can be dif-ferentiated, de-multiplexed and switched individually withinthe actual network layer is treated as a Label Switched Path(LSP) in GMPLS, like for instance a time slot in a time-divi-sion multiplexed infrastructure or a wavelength channel onan optical fiber.

On the one hand, the abstraction of LSPs makes it possi-ble to monitor and control the entire stack of network lay-ers by a common control infrastructure, thus advancing theconvergence of new and legacy technologies and the seam-less interconnection of heterogeneous networks. From thestandpoint of routing and Traffic Engineering, on the otherhand, the integrated view of the network (the so called Peermodel) implies that the routing entity has combined resource

123

Photon Netw Commun

and topology information from all the network layers, whichfacilitates for attaining better network efficiency than it ispossible under the strict separation of control functionalities,enforced by the conventional overlay model.

It is a design decision made early in the course of definingthe GMPLS control plane that the standards suite does notspecify explicitly the exact routing algorithm to be used toset up LSPs. The GMPLS standard only describes the envi-ronment, the functional model, and the mode of operation ofa hypothetical GMPLS routing algorithm. Accordingly, in aGMPLS-based multi-layer network architecture the task ofthe routing algorithm can be posed as follows (the so calledConstraint-based routing problem): given the virtual graphrepresenting the physical network infrastructure, the alreadyestablished lower-layer LSPs subject to grooming and theswitching capability of the network nodes, find a path fora new LSP request from the given source interface to thegiven destination interface, subject to a number of opera-tional constraints like, e.g., the type of applicable protection,required bandwidth, etc. Observe that the constraint-basedrouting problem in multi-layer networks is more complexthan in traditional, single layer networks, since it is not con-fined to the conventional task of finding an appropriate for-warding path that fulfills the constraints, but now it is alsoup to the routing entity to decide in which layer to serve therequest. Since the network consists of two or more techno-logical layers stacked on top of each other, and any tech-nological layer is eligible to host a new request, the routingalgorithm has to decide whether to set up a new lower layerLSP, which later can be subjected to host further upper layerLSPs, or to groom the request into a sequence of alreadyexisting LSPs (or a mixture of the two options). As shall beshown, the rule for choosing the right layer to set up new LSPs(the so called layer-preference policy) has crucial impact onthe performance of the network. Therefore, studying andevaluating layer-preference policies, within the technolog-ical context defined by the GMPLS paradigm, stands in thefocus of main interest in this paper.

After a quick survey of the literature on layer-preference(Sect. 2), we define two simple layer-preference policies andwe discuss the respective pros and cons. We find that nei-ther of these simple policies is adequate to govern layer-preference as they cause adverse performance degradationin the long-run. Therefore, in Sect. 3 we propose a novelheuristic called the Min-phys-hop routing and wavelengthassignment algorithm. We argue that this algorithm is notonly efficient but also highly practical, because it readilylends itself to be implemented in contemporary GMPLS net-works. Consequently, in Sect. 4 we discuss the deployabilityof the Min-phys-hop algorithm, taking into account the con-ventions imposed by the GMPLS standards and the techno-logical restrictions imposed by operational network devicesof our days. To further stress that the proposed algorithm

is really viable in practice, in Sect. 4.1 we sketch two refer-ence scenarios offering a seamless deployment path towards afull-fledged GMPLS-enabled network architecture. The finalpart of the paper, Sect. 5, is devoted to evaluate and comparelayer-preference policies. We define a unified framework forthis purpose and we present the results of comprehensivesimulation studies, which provide evidence that layer-pref-erence is a critical factor in the emergent performance of amulti-layer network and that the Min-phys-hop algorithm isfairly efficient in this regard. Finally, we conclude the paperin Sect. 6.

The GMPLS framework is designed for massively multi-layer networks. Wherever possible, such general networkswith more than two layers will be considered throughoutthe paper. However, since the most popular setup containsonly two layers, namely an IP/MPLS layer on top of a(Dense) Wavelength-Division Multiplexing (DWDM) opti-cal infrastructure, sometimes we shall restrict ourselves tothis very two-layer setup. When not stated otherwise, thelower layer we shall call the optical layer, its LSPs we shallcall lightpaths, and the term LSP will be usually meant todenote IP/MPLS connections.

2 Backgrounds

The GMPLS architecture is described in large detail in[1–3]. The signaling framework can be found in [4] while theroutingmodel isdescribed in [5]. Implementation-specificco-nsiderations can be found in [6] (particularly concerning, theOpenShortestPathFirst routingprotocol-TrafficEngineeringextensions [OSPF-TE] as described in [7]) and [8] (the samefor IS-IS-TE, the Intermediate-System-to-Intermediate-Sys-tem routing protocol Traffic Engineering extensions).

A good introductory material on routing and wavelengthassignment algorithms can be found in [9], and multilayertraffic engineering (MLTE) is reviewed in [10]. DynamicMLTE schemes are a well-researched area, for a good intro-duction the reader is referred to [11,12]. It must be noted,however, that specifically layer-preference policies, the mainquestion we investigate here, has never been explicitlyaddressed in the literature, this problem only gets some mar-ginal treatment. In particular, in [13–15] so called groom-ing policies are identified, which govern the way a layer isselected to host a new LSP. A grooming policy, for instance,would be defined as “insert the LSP to the direct lightpathfrom the source to the destination if one is available, other-wise establish a new direct lightpath, and if both attempts fail,use a combination of existing and new lightpaths”. The prob-lem with grooming policies is that they are rigid in the sensethat they are unable to express mixed layer-preference poli-cies, taking into account the implied resource consumption(like “set up a new lightpath of length at most two hops, and

123

Photon Netw Commun

if this attempt fails, use a mixture of existing and new light-paths”). An early attempt to define such mixed-strategies isdescribed in [16]. The most important findings of these worksare that although certain layer-preference policies work wellunder specific circumstances (like in a lightly loaded net-work or one with unlimited wavelength conversion), theredoes not seem to exist a universally optimal static policy.This paves the way for a mixed layer-preference policy, asshall be discussed in the next section.

3 Layer-preference policies

First, we demonstrate the problem of constraint-based rout-ing in multi-layer networks and the importance of layer-preference through a simple example. Consider the physicalnetwork topology depicted in Fig. 1. There are four com-bined IP/(D)WDM router devices (1, 2, 4 and 5) and anOptical-Cross Connect (OXC) node (node 3) in the network.These devices have different switching capabilities: whilean OXC device can only switch between optical links, anIP/(D)WDM device can initiate and terminate lightpaths andis able to switch individual wavelengths as well (throughmultiplexing/demultiplexing wavelength channels throughthe IP layer). Suppose that, at the point in time when welook at the network, two lightpaths have already been estab-lished: LP1 from node 2 to node 5 (via OXC 3) and LP2 fromnode 5 to node 4 (again via OXC 3), both offering enoughcapacity to admit a new request. Further, suppose that a newGMPLS connection request has to be routed between nodes1 and 4.

In the traditional overlay model, the IP/MPLS layer (theclient layer) requests a new direct lightpath between thesource and the destination node from the (D)WDM layer(the server layer) and the connection request is tunneled intothe new lightpath. However, allocating a new lightpath is a

1 2 3 4

5LP1 LP2

Fig. 1 Virtual graph representation of a sample network scenario.There are two wavelength channels at each link: fat pipes representoptical links and thin pipes within optical links are wavelength chan-nels. There are two lightpaths, LP1 and LP2, both consuming a singlewavelength channel on intermediate optical links

tedious task and it may very well be impossible, if any of theintermediate nodes run out of spare wavelength channels. Inour example, we can only instantiate one connection betweennode 1 and 4 under the overlay model, since the first con-nection consumes the remaining spare wavelength channelsbetween node 2 and 4, and there is no room to accommodatethe second request along a direct lightpath.

In contrast to the overlay model, the peer model allowsfor routing a request to a path that is obtained as the con-catenation of new and already existing lightpaths. That is, ifno direct lightpath can be established we can still route theconnection through a newly created lightpath between nodes1 and 2, plus the concatenation of LP1 and LP2. At inter-mediate nodes, the GMPLS LSPs are demultiplexed, and thepackets of the new request are groomed into the next light-path. The additional flexibility offered by the peer model overthe overlay model usually leads to substantial boost in theprofitability and the useful throughput of the optical network[13]. Therefore, we mostly concentrate on the peer modelhenceforth.

In a GMPLS network using the peer model, the routingentity has integrated knowledge on the configuration andresource availability of all the technological layers that makeup the protocol stack, and therefore it has the authority toselect the layer at which new requests will be routed at. Wecall the set of rules governing the choice of the preferredlayer in which to accommodate route requests as the layer-preference policy.

Given an incoming request, the first task a routingalgorithm performs is to check whether there exists a directlightpath with sufficient capacity between the source and thedestination. If such direct lightpath exists, then the new LSPis inserted into this lightpath right away. If no such lightpathexists, however, then one can choose between two simple,contradictory layer-preference policies: routing in the lowerlayer and routing in the upper layer. The first policy favorssetting up direct low-level LSPs for incoming requests, whilethe latter one prefers grooming upper layer LSPs into lowerlayer LSPs whenever possible.

In the rest of this section, we discuss these layer-preferencepolicies in more detail. We will show that there is not a singleone-fits-all solution and that favoring any single layer of theIP/(D)WDM architecture comes with its very own specialadvantages and drawbacks. Then we define a new heuris-tics, the Min-phys-hop algorithm, aimed at overcoming thelimitations of simple layer-preference policies.

3.1 Routing in the lower layer

One obvious choice is to push routing into the lower layer,that is, to serve a new request in the bottommost layer inthe stack that can handle it. In a two-layer IP/(D)WDM net-work, this would always amount to instantiate a new direct

123

Photon Netw Commun

lightpath for a new LSP request, and only attempt to reuseexisting lightpaths once setting up a new one fails due to thelack of appropriate resources at the optical layer.

Unfortunately, routing in the lower layer causes the fre-quent setting up of lower-layer LSPs, which, by nature, tendto have an abundance of capacity but are tedious to estab-lish and tear down, and might consume expensive resourcesdue to frequent Optical-Electronic-Optical (OEO) transfor-mations along the path. More regrettably, however, this pol-icy causes an adverse saturation phenomenon: as the networkis filled up with traffic from different source and destinationnodes, a huge number of direct but hardly ever used lightpathswill be established. At some point the network runs out ofspare wavelength channels, and there remains no other choiceto accommodate a new LSP request than to use a lengthycombination of existing lightpaths (since no new ones can bebuilt), which will certainly cause suboptimal routing in thelong run. Since this phenomenon shows exceptionally strongresemblance to the fragmentation of memory and disk blocksin computers, we call it wavelength fragmentation [17]. Fora demonstration of wavelength fragmentation, see the simu-lation results in Section 5.3.

3.2 Routing in the upper layer

A way to avoid wavelength fragmentation is to push routinginto the uppermost layer possible, that is, to reuse existinglower-layer LSPs to host new upper-layer LSPs as long as itis possible, and only apply to lower layers when it is abso-lutely unavoidable. The problem here is that a lower-layerLSP, represented as a direct link in the virtual graph (a socalled TE-link in the GMPLS terminology), does not offerany tangible information for the routing algorithm as to howmuch real physical resource it uses and what does it cost (interms of optical transmitters/receivers, electronic resources,etc.) to groom the new LSP into it. This often tricks the tradi-tional shortest path routing algorithm to choose exceedinglylong and costly paths, which, when viewed from the physicallayer, may even contain physical level loops.

To understand the emergence of loops, consider the sam-ple scenario depicted in Fig. 1. Here, routing in the upperlayer yields that for a request from node 2 to node 4, thesequence of LP1 and LP2 will be assigned as a forwardingpath. However, we observe that the path in the optical layertaken by the packets of this connection contains a loop, sincenode 3 will be hit twice, once on the way through LP1 and yetanother time on the way through LP2. One of the most impor-tant results of this paper is the revelation that such routingloops pose a significant obstacle in optimizing a peer-archi-tecture. We shall point out that routing loops have uniquelyadverse effect on the routing performance, even though, asshall be shown below, routing loops are a natural concomi-

tant phenomenon inherent to the peer model (and, in somecases, to the overlay model as well).

Theorem 1 Consider a peer model virtual topology, in whichnodes represent switching devices and links represent eitherwavelength channels or lower-layer LSPs in the form of TElinks. Now, the decision question whether there exists a paththat does not contain loop(s) at the lower layer is NP-com-plete.

Proof The transformation is straightforward from the Pathwith Forbidden Pairs problem (GT54) [18]. ��

3.3 The Min-phys-hop algorithm

As it turns out, it is completely hopeless to devise a routingalgorithm that can always avoid creating loops. Instead, onemust resort to viable heuristics. Therefore, we developed anovel heuristic, which we call the Min-phys-hop algorithm.The heuristic is based on the idea that in order to a pathto be efficient, it should traverse as few physical nodes aspossible. For this, we label each link in the virtual graphthat describes the integrated knowledge on the network layerstack by the physical length of the LSP it represents, andchoose the least-cost path in the resultant weighted graph.This way, short, direct lightpaths will always be preferredover exceedingly long LSPs and loops are avoided as longas possible. For a formal description of the Min-phys-hopalgorithm specialized to two-layer IP-MPLS/(D)WDM net-works, see Fig. 2.

Consider the network scenario depicted in Fig. 1 and sup-pose that the task is (again) to set up an LSP from node 2 tonode 4. Now, using the Min-phys-hop algorithm amounts toset the weight of both LP1 and LP2 to 2 (that is, the numberof optical links the lightpath traverses), the weight of all theother links to 1 and find the minimum cost path. This meansthat, in this specific example, a new direct lightpath will beset up from node 2 to 4, which means that no routing loopswill be created in physical layer.

Fig. 2 The Min-phys-hop routing algorithm

123

Photon Netw Commun

4 Deploying the Min-phys-hop routing algorithmin GMPLS networks

The purpose of the Min-phys-hop algorithm is to performrouting and wavelength assignment in multi-layer networksand govern LSP grooming. As such, for it to be really usefulin practice, it must fit perfectly into the GMPLS framework.In this section, we show that the Min-phys-hop algorithmlends itself readily to be implemented and deployed in con-temporary GMPLS networks.

The GMPLS framework is a remarkably feature-rich one,embracing a vast number of different network technologies,routing models and modes of operation. Before evaluatingthe applicability of the Min-phys-hop algorithm for GMPLS,we need to review a number of important technological ques-tions.

Although the GMPLS paradigm has been extendedrecently to be able to handle inter-domain LSPs spanningmultiple Autonomous Systems (ASs) [19], below we onlyconcentrate on an intra-domain scenario. We shall assumethat the routing entity holds complete and (relatively) up-to-date information on the topology and resource availability inits local AS. This assumption is in line with the rest of theliterature and the present state-of-the-art in GMPLS technol-ogy.

The two most important routing protocol infrastructuresof the GMPLS protocol suite are OSPF-TE-GMPLS (theGMPLS extensions to the OSPF-TE [6]) andIS-IS-TE-GMPLS (The GMPLS extensions to the IS-IS-TE[8]). Since the functionality provided by these protocols ismore or less identical from the viewpoint of GMPLS, weshall consider the Min-phys-hop algorithm only in terms ofOSPF-TE-GMPLS. Naturally, all of our findings are equallyvalid to IS-IS-TE-GMPLS as well.

The GMPLS standards do not specify the exact locationof the routing entity within the network: routing might bedistributed amongst cooperating interior gateway protocol(IGP) entities throughout the network, or it might be cen-tralized in the so called Path Computation Elements (PCE,[20]) located anywhere within, or even outside the domain.Below, we shall deal with both of these scenarios. Next, weoverview the questions, which help us to answer to how theMin-phys-hop algorithm fits into the GMPLS framework.

The first question we ask is whether the mode of operationof the algorithm fits into that of GMPLS. In the usual con-text of constraint-based routing, LSP setup requests arriveone-by-one at the routing entity, which then carries out cal-culations to find an appropriate path, subject to constraintsincluded in the request. This mode of operation is calledon-demand routing, and it is the default mode of GMPLSrouting. The opposite is route precomputation: here, pathsare precomputed for all possible route requests, and incom-ing requests are served from this precomputed routing table.

This is the basic mode of operation of IP networks, and it isexpected that some form of route precomputation will find itsway into GMPLS networks as well (e.g., to serve the upper-most IP layer). While the Min-phys-hop algorithm perfectlyserves the needs of on-demand routing, it is still important toinvestigate whether it supports precomputation as well, andif yes, then to what extent.

The answer is generally yes, though with limitations.Under the hood, the Min-phys-hop algorithm is nothing morethan labeling the edges with the physical length of the under-lying objects and performing shortest path computations overthe resultant graph. But shortest path algorithms have for longmanifested an obvious choice for route precomputation, sofor the first sight there does not seem to be any difficultyhere. The problem is that in architectures following the peermodel, where the entire stack of all network layers is exposedto the routing algorithm, it is allowed to initiate lower layerLSP setups upon servicing an upper layer LSP request. How-ever, this might change the topology of the virtual graph(e.g., in an IP-MPLS over (D)WDM setup, when a lightpathis established, the corresponding wavelength edges shouldbe dropped from the virtual graph), and there is no way tomake this change visible to other LSP requests being underprecomputation. Therefore, the Min-phys-hop algorithm isonly usable for precomputation when the lower layers arenot allowed to change during the calculation of the routingtable. Such a setup basically accounts for an overlay-mod-eled network, where there is no integration and sharing ofrouting information between the layers. As a summary, wecan state that the Min-phys-hop algorithm is only usable forroute-precomputation in an overlay-based architecture.

For the Min-phys-hop algorithm to be usable in the con-text of GMPLS, it is essential that the underlying routingprotocol machinery, the GMPLS extensions to the OSPF-TErouting protocol in our case, make all the data available that isnecessary to execute the algorithm. The most important dataneeded for Min-phys-hop is—apart from the virtual graphconstructed from the Traffic Engineering Database (TED)describing the network and the switching capabilities of net-work nodes—the length of the lower-layer LSPs in terms ofthe number of physical hops they traverse.

Unfortunately, as of the present state-of-the-art, neitherOSPF-TE nor the GMPLS extensions include the informa-tion on the physical length of TE-links in the Link StateAdvertisements (LSAs) generated to describe these elements.OSPF-TE [7] adds the following set of data to the onesdefined in the original OSPF standard to describe TE-links:

– Link type: the type of the link, either point-to-point ormulti-access

– Link ID: to uniquely identify the other end of the link– Local interface IP address: the IP address of the local

interface corresponding to this link

123

Photon Netw Commun

– Remote interface IP address: the IP address of the neigh-bor’s interface corresponding to this link

– Traffic engineering metric: link metric for traffic engi-neering purposes; different than the standard OSPF linkmetric and assigned by a network administrator

– Maximum bandwidth: maximum bandwidth that can beused on this link

– Maximum reservable bandwidth: maximum bandwidththat may be reserved on this link; may be greater thanthe maximum bandwidth in which case the link may beoversubscribed

– Unreserved bandwidth: amount of bandwidth not yetreserved on this link

– Administrative group: bit mask assigned correspondingto the administrative group (Class or Color) assigned tothe interface

Additionally, the GMPLS extensions for OSPF-TE stan-dard adds some further enhancements to the TE propertiesof GMPLS TE links. Encoding of this information in OSPFis specified in [6]:

– Support for unnumbered links: unique link identifier if thecorresponding interfaces do not have separate IP addresses

– Link Protection Type: protection capability for the link(Unprotected, 1+1, 1:1, etc.)

– Shared Risk Link Group Information: unique SRLG iden-tifier(s) describing the SRLG(s) the link belongs to

– Interface Switching Capability Descriptor: to identify theswitching, multiplexing and de-multiplexing capabilitiesof the interfaces connected to the link

Unfortunately, the physical length of the TE-links is absentfrom the set of properties used to describe a lower-layer LSPin both OSPF-TE and OSPF-TE-GMPLS. Hence, there is nostraightforward way to encode this information into the vir-tual graph and thus the Min-phys-hop algorithm has no waysto differentiate between the resource usage of TE-links. Theonly possibility is to allocate the OSPF-TE property “Trafficengineering metric” to this purpose. That is, GMPLS LabelSwitch Routers (LSRs) that originate or terminate a LSPencode the physical length of the LSP in the “Traffic engi-neering metric” of the LSA generated to describe that LSP.This LSA is then appropriately flooded throughout the net-work by OSPF-TE, conveying the required information to allLSRs in the domain.

The unique purpose of the Min-phys-hop routing algo-rithm is to select paths so that the induced usage of the valu-able network resources are minimized. This is reflected (asevidenced by the simulation studies discussed later) in thereduction on the number of physical-level loops and in theaverage length of the paths. However, in a realistic network

setting there might arise further requirements and constraintsimposed on the returned path, other than simplistic minimi-zation of network resources, including:

– Minimum bandwidth: all links of the path should offer atleast the specified amount of bandwidth in the “Unreservedbandwidth” link descriptor

– Maximum acceptable delay– Class or color: restrict the path to a specific administrative

class of links– Minimum acceptable protection: restrict the path to exclu-

sively to e.g., 1+1 or 1:1 protected links– Adaptation: LSPs of specific adaptations and payload

structures can be requested, like, for example, a VC-3 Syn-chronous Digital Hierarchy (SDH) circuit

– Interface switching capability: since a GMPLS networkmight span various network layers, it is possible to con-fine the selected path into a particular network layer

Since any of these constraints might be rightfully imposedeither in itself or combined with some other one, it is essen-tial to review how the Min-phys-hop algorithm can handleconstraints on the selected paths and how it mixes with tra-ditional constraint-based routing algorithms.

There are in essence two approaches to constraint-basedrouting. There exist certain constraints that can be satisfied aseasily as filtering the links in the virtual graph on which pathselection is carried out. For instance, finding a path fulfillinga certain minimum bandwidth requirement can be done bysimply removing all links of capacity lower than the require-ment from the virtual graph and returning any paths in thepruned virtual graph. Not just that these bottleneck type ofconstraints are easy to handle, but they also mix quite well(meaning that it is straightforward to fulfill two or more bot-tleneck type constraints at the same time: just filter all thelinks violating any one of the imposed constraints). Unfortu-nately, additive type of constraints (like e.g., delay or admin-istrative cost) are much harder to consider. These constraintsare called additive because the quantity describing a particu-lar path equals the sum of the quantities describing its links.The problem is that additive type of constraints do not mixwell: selecting a path subject to two or more additive typeconstraints at the same time is NP-hard. This means that theMin-phys-hop algorithm (which involves its very own addi-tive metric in the constraint-based routing calculation: thephysical length of the TE-links) is not suitable to computedelay-constrained paths, because the number of additive typeof metrics to be considered would be two, rendering the pathselection problem NP-hard. On the other hand, practicallyany of the remaining constraints are easy to incorporate intothe Min-phys-hop algorithm, like minimum bandwidth, min-imum protection type, etc., since these are all bottleneck typeconstraints.

123

Photon Netw Commun

Next, it is important to examine, how the Min-phys-hopalgorithm supports networks consisting of more than two lay-ers stacked on top of each other, like e.g., a Packet-SwitchCapable layer (e.g., MPLS) on top of a Time-Division-Multi-plex Capable layer (e.g., SDH) on top of a Lambda-SwitchCapable layer (e.g., (D)WDM). This is permitted and, toa large extent, fostered by the GMPLS framework. Thus,the question naturally emerges: how does the Min-phys-hopalgorithm handle LSP-hierarchies as described in [21]? Sincethe metric defined by the physical length parameter is stack-able—that is, the physical length of a higher-layer LSP is thesum of the physical length of the lower-layer LSPs it consistsof, which are again labeled by the sum of the still-lower-layerLSPs—, the Min-phys-hop algorithm correctly generalizesto GMPLS networks incorporating more than two switch-ing capabilities. Note that, however, we do not address thisscenario in our simulations.

Finally, it is important to examine how the Min-phys-hopalgorithm performs in real GMPLS networks. Since we donot have an appropriate-sized GMPLS test bed at our dis-posal to test the algorithm on, we need to confine ourselvesto simulation studies. This is the main topic of the rest ofthis paper, but first we sketch some likely scenarios in whichthe Min-phys-hop algorithm may find its use in GMPLS net-works.

4.1 Scenarios for deployment

After comprehensive evaluations, it seems that the Min-phys-hop algorithm readily fits into the GMPLS framework. It onlyuses routing information that is made available by OSPF-TE-GMPLS to it, it supports multiple layers, LSP hierarchies andboth on-demand routing and precomputation. Consequently,it seems plausible to consider deploying it in GMPLS net-works. Below, we sketch two potential deployment scenarios.

The most likely deployment path towards GMPLS standsin the gradual upgrading of today’s IP-MPLS over (D)WDM

networks towards a complete GMPLS stack by introduc-ing the (D)WDM layer to under the authority of the unifiedGMPLS control plane [22]. As the first step of this process,it is expected that an overlay-modeled control architectureis implemented instead of a full-fledged peer architecture.In such a network architecture (see Fig. 3a), routing is dis-tributed among the IGP entities residing on LSRs across therouting domain. These OSPF-TE-GMPLS protocol entitiessee an overlay model of the network stack (in which light-paths from the (D)WDM layer are represented as TE-links butadditional (D)WDM network layer infrastructure is invisible)and, based on this virtual graph, precompute a full SPF treeto all IP prefixes available in the domain using the Min-phys-hop algorithm. Once there is no path available to a destinationprefix (either because there is no connectivity to the prefix orthe capacity at the corresponding lightpaths is exhausted), themanagement plane solicits the routing entity responsible forthe (D)WDM layer to establish a new lightpath towards theseprefixes. These requests are served by the conventional rout-ing machinery of the (D)WDM layer. In this model, integrat-ing the legacy networking technologies into GMPLS is doneonly half-way: the forwarding planes are handled commonlyvia the notion of abstract GMPLS labels (so the label spaceis shared), but the routing functionality is still unshared. Wesee that the Min-phys-hop algorithm fits perfectly into sucha scenario, although more than two layers, constraint-basedrouting and full-fledged TE is generally not available.

As the next, long term aim of GMPLS, not just theforwarding plane functionality but routing control as wellis expected to become unified. This is possible by buildinga complete peer model of the entire GMPLS technologicalstack in which every piece of network element, interface,switching capability and forwarding channel is represented(see Fig. 3b). Since building and managing such an expen-sive network representation might impose too much burdenon the LSRs, a PCE can be installed in the network to carryout routing calculations on behalf of the LSRs in the domain

(a) (b)

Fig. 3 Deployment scenarios for the Min-phys-hop algorithm in IP-MPLS/(D)WDM networks

123

Photon Netw Commun

[20]. In a PCE-based architecture, an LSR willing to set upa new LSP makes a routing request to the PCE responsi-ble for the domain. The PCE learns routing information (forinstance, through participating in the flooding process of theIGP), builds a TED from the collected information and con-structs a graph describing the entire network stack, in whichall elements are marked with the amount of free capacityavailable at that element. Additionally, for the purposes ofexecuting the Min-phys-hop algorithm, TE-links are labeledwith the physical length of the underlying LSP or the aggre-gate length of the underlying LSP hierarchy. This quantityis made available by OSPF-TE-GMPLS in the “Traffic engi-neering metric” TE-link description attribute. Using the vir-tual graph, the PCE computes a suitable path subject to theconstraints communicated by the initiating LSR to the PCEand returns that path to the LSR. It is also possible to keepthe LSR and the PCE synchronized, either to inform the lat-ter of the success or the failure of setting up the LSP or tonotify the former on the availability of better paths that mighthave become usable meanwhile. A PCE based architecture isadvantageous, not just because it helps relieve LSRs from theburden of routing, but also because the dedicated hardware ofthe PCE makes it possible to invoke more sophisticated andmore complex routing algorithms, executed in an on-demandfashion. Again, the Min-phys-hop algorithm seems a perfectmatch to perform constraint-based routing at the PCE.

5 Simulation studies

In the previous section, we concluded that the Min-phys-hopalgorithm lends itself readily to be deployed in GMPLS net-works. The motivations for doing so are multi-faceted: thealgorithm is simple and easy to implement since it is, underthe hood, a simple least-cost path algorithm with special linkweights. Additionally, the original motivation to develop theMin-phys-hop algorithm was to find a trade-off between twocontradictory routing policies on choosing the appropriatelayer to host LSP requests, namely, whether to push routingtowards the uppermost technological layers in the GMPLSstack or rather to serve requests in the bottommost layer pos-sible. Finding such a trade-off is important, because we foundthat both of these policies cause adverse phenomena, namelywavelength fragmentation and/or routing loops.

In this section, we support these claims by empirical evi-dence: we present the results of extensive simulations todemonstrate how those adverse phenomena emerge and that,by using the Min-phys-hop algorithm, we can avoid thesepitfalls and generally achieve better network performance.First, we show a simple means to model limited wavelengthand optical-electrical-optical (OEO) conversion capabilitiesin the simulated networks and we discuss a way to representlayer-preference policies, designed to capture all the impor-

tant characteristics of layer-preference policies discussed ear-lier in this paper. Then, we turn to discuss the simulationresults.

5.1 Optical device model

In our simulation studies, we concentrated on two-layer IP-MPLS/(D)WDM networks. The main constituent of such anetwork is an OXC device. Figure 4 depicts the structure of atypical N ×N OXC of our days. It has N input and N outputports, S wavelengths at each incoming and outgoing fiber andit can switch any particular λi wavelength from any incomingport to the same λi wavelength on any outgoing port. Addi-tionally, this OXC can drop (and add) exactly S channels byintroducing the corresponding wavelengths to optical receiv-ers (transmitters) for further electronic processing. Electronicprocessing is also the way for wavelength conversion in thisdevice, that is, there is no optical domain wavelength con-version available in the OXC. Also note that a certain λi

wavelength can be dropped from exactly one incoming portand it is not possible to drop the same λi wavelength from twoor more incoming ports at the same time. The same appliesto adding wavelengths to outgoing ports. This restriction willbe important, because our model for the OXC, described indetail in the sequel, is designed deliberately to reflect thistype of interior contention of today’s OXC devices.

The model we used to represent limited OEO conversioncapability is depicted in Fig. 5a. There are N input and Noutput ports, however, since our graph model is in essenceundirected, we did not differentiate between incoming andoutgoing interfaces. Additionally, all the S wavelengths at theconnected fibers are represented by individual wavelength

Fig. 4 Typical optical crossconnect (OXC) device

123

Photon Netw Commun

(a) (b) (c)

Fig. 5 OXC device model with three input–output ports and three wavelength channels (N = 3, S = 3)

edges of capacity CWL. The electronic point, which corre-sponds to the “Add Drop branch” in Fig. 4, is represented bythe point E and the optical receivers (OE conversion) andoptical transmitters (EO conversion) are modeled by capaci-tated edges from the wavelength edges to the electronic point.The capacity equals M×CWL, where M manifests restrictedOEO conversion capability. For M = 0 there is no electroniclayer and no OEO conversion, and setting M to infinity meansthat there is unlimited OEO conversion.

For M = 1, the model accurately reflects the OXC deviceof Fig. 4 with all its capabilities and limitations. More spe-cifically, our model correctly encodes the restriction that oneparticular λi can only be dropped (and added) from just oneincoming port (to one outgoing port) at the same time. Notealso that our model is remarkably flexible in the sense that it isable to express many more optical switching equipments, notjust the OXC device above. In particular, Fig. 5b shows themodel of an OXC device without electronic layer (that is, the“Add Drop branch” is absent from the device) and Fig. 5cdepicts the model for an OXC with unlimited wavelengthconversion capability in the optical domain.

5.2 Representing layer-preference policies

In order to model different layer-preference policies in oursimulations, we transform these policies into special rules forsetting link weights in the virtual graph. First, we build thevirtual graph representation of the integrated IP-MPLS/(D)WDM network topology [23] using the extended OXC-modelabove and then we use Dijkstra’s shortest path routing algo-rithm to compute paths over specially assigned link weights.In particular, the weights of the lightpath links (ωLP)—theselinks stand for already established lightpaths—and wave-length links (ωOP) is chosen as 1, and the weight of the restof the links is set to a very small positive constant. Then, thelayer-preference policy is manifested in the course of path

selection by rescaling the link weights by a configurable α

parameter as follows:

ωLP ← 1

αωLP

ωOP ← 1

1− αωOP

Observe that setting α = 0 pushes routing into the lower(D)WDM layer, because the weight of lightpath links is setto infinity in this case. Contrariwise, setting α = 1 yieldsthat it is cheaper to accommodate a new LSP on a series ofalready established lightpaths. Moreover, all other settings ofα between 0 and 1 represent different trade-offs between thetwo layer-preference policies, which is not possible withinprevious models [13–16]. Finally, there remains to imple-ment the Min-phys-hop algorithm in our simulations, butthis is easy: simply let ωOP = 1 and set ωLP to the numberof hops the corresponding LSP traverses (see Fig. 2).

5.3 Evaluation of simple layer-preference policies

Below, we demonstrate the adverse phenomena, namelywavelength fragmentation and routing loops, caused bysimple layer-preference policies. We also include the Min-phys-hop algorithm in the simulation results to show that thisalgorithm is immune to such phenomena.

In the course of our simulations, we filled up the KL-net-work [23] with long-lived connection setup requests one-by-one and observed the evolution of the network. The requestsource and destination pairs were selected according to inde-pendent uniform distributions and the request size was uni-formly 1 unit. All optical edges could carry a total numberof three wavelengths, and each such channel was of 10 unitsof capacity. For the sake of simplicity, nodes were of unlim-ited OEO conversion capability. We executed 500 simulationrounds, which was enough to saturate the network.

123

Photon Netw Commun

Figure 6 shows the number of failed connection requestsas the function of α at the end of the simulation. The mostimportant observation is that routing in the upper layer(α = 1) is inferior to routing in the lower layer in our simula-tions, as it blocks almost twice as many connection requests.This is clearly attributed to the fact that routing in the upperlayer is exceptionally prone to creating routing loops, whichyields overly long paths and hence the network gets saturatedprematurely. The results in Fig. 7 (showing at every timeinstance the average length of the last 20 selected paths) andFig. 8 (depicting the number of physical-level loops as thefunction of time) further substantiate this observation.

It seems that, in this specific example, it is much more ben-eficial to encourage routing in the optical layer. The lowerthe value of α the lower the average length of the paths andthe number of routing loops all the way to the point that withα = 0, no loops are created until the 220th time step. How-ever, at this point the vast majority of free wavelengths is usedup and a dramatic increase in the path length and the numberof loops can be observed. This is because only (often very

30

35

40

45

50

55

60

0 0.3 0.4 0.5 0.6 0.7 1 min-phys-hopalpha

Nu

mb

er o

f b

lock

ed r

equ

ests

Fig. 6 Number of failed connections for difference setting of α andthe Min-phys-hop algorithm

0.5

1

1.5

2

2.5

3

36 72 108 144 180 216 252 288 324 360 396 432 468Time [s]

Avg

. len

gth

of

the

last

20

pat

hs

00.30.51min-phys-hop

Fig. 7 Average length of the last 20 paths at every time instance, fordifferent settings of α and the Min-phys-hop algorithm

00.30.51min-phys-hop

0

2

4

6

8

10

12

14

16

36 72 108 144 180 216 252 288 324 360 396 432 468

Time [s]

Nu

mb

er o

f lo

op

s

Fig. 8 Number of loops in the physical network for different settingsof α and the Min-phys-hop algorithm

long) paths through existing lightpaths remain available. Thisis the phenomenon we called as wavelength fragmentationin Sect. 3.1. We note that after some peak value the length ofthe paths begins to drop, because, due to the fragmentation,no paths of continuous wavelength edges can be found andonly requests between adjacent nodes succeed.

We observe that the Min-phys-hop algorithm implementsa decent trade-off between the two policies. Its reluctanceto avoid forming physical-level loops remains unparalleledby the other routing policies. Only when there is practicallyno loop-free path over the physical network it selects loopypaths. Simultaneously, it creates short paths, which yieldsfewer blocked calls.

One might argue that our observations are specific to theselected configuration. Therefore, we repeated the simula-tions for different settings of the number of wavelengthsand their capacity (ensuring that the product always remains30), though, the adverse effects of wavelength fragmentationand routing loops were always clearly manifested. Figure 9shows the number of blocked calls at the end of the simula-tions in grand comparison. We observe that if there is only

150

200

250

300

350

400

0 0.3 0.4 0.5 0.6 0.7 1 min-phys-hopalpha

Nu

mb

er o

f b

lock

ed r

equ

ests

1 wavelength2 wavelengths3 wavelengths5 wavelengths10 wavelengths15 wavelengths

Fig. 9 Number of failed connections under different configurations ofwavelength number and wavelength capacity for different settings of α

and the Min-phys-hop algorithm

123

Photon Netw Commun

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0.05 0.1

0.15 0.2

0.25 0.3

0.35 0.4

0.45 0.5

0.55 0.6

0.65 0.7

0.75 0.8

0.85 0.9

0.95

Alpha

Ave

rag

e C

all B

lock

ing

Rat

io

Min-hop4Min-phys-hop4Min-hop8Min-phys-hop8

0

0.02

0.04

0.06

0.08

0.1

0.12

0.05 0.1

0.15 0.2

0.25 0.3

0.35 0.4

0.45 0.5

0.55 0.6

0.65 0.7

0.75 0.8

0.85 0.9

0.95

Alpha

Ave

rag

e C

all B

lock

ing

Rat

io

Min-hop16Min-phys-hop16Min-hop32Min-phys-hop32

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0.05 0.1

0.15 0.2

0.25 0.3

0.35 0.4

0.45 0.5

0.55 0.6

0.65 0.7

0.75 0.8

0.85 0.9

0.95

Alpha

Ave

rag

e C

all B

lock

ing

Rat

io

Min-hop4Min-phys-hop4Min-hop8Min-phys-hop8

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0.05 0.1

0.15 0.2

0.25 0.3

0.35 0.4

0.45 0.5

0.55 0.6

0.65 0.7

0.75 0.8

0.85 0.9

0.95

Alpha

Ave

rag

e C

all B

lock

ing

Rat

io

Min-hop16Min-phys-hop16Min-hop32Min-phys-hop32

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

0.05 0.1

0.15 0.2

0.25 0.3

0.35 0.4

0.45 0.5

0.55 0.6

0.65 0.7

0.75 0.8

0.85 0.9

0.95

Alpha

Ave

rag

e C

all B

lock

ing

Rat

io

Min-hop4Min-phys-hop4Min-hop8Min-phys-hop8

0

0.05

0.1

0.15

0.2

0.25

0.05 0.1

0.15 0.2

0.25 0.3

0.35 0.4

0.45 0.5

0.55 0.6

0.65 0.7

0.75 0.8

0.85 0.9

0.95

Alpha

Ave

rag

e C

all B

lock

ing

Rat

io

Min-hop16Min-phys-hop16Min-hop32Min-phys-hop32

(a) (b)

(c) (d)

(e) (f)

Fig. 10 Average call blocking ratio (CBR) produced by differentlayer-preference policies and the Min-phys-hop algorithm for variousnumber of wavelengths per optical link, for networks of unlimited wave-length conversion (a), (b) and limited wavelength conversion (c), (d)

(for M = 1) and (e), (f) (for M = 2). The number after the name of thepath selection mechanisms represents the number of wavelengths peroptical link in the network

one (but of relatively large capacity) wavelength channel ateach link, then it becomes notoriously hard to optimize thenetwork. This is because the virtual topology is adjusted in

extremely large steps in this case. Therefore, one might preferto route the requests in the upper layer and avoid the inter-vention at the optical layer as much as possible. However, as

123

Photon Netw Commun

the number of wavelength channels decreases so the adverseeffects of routing loops increases, and routing in the opticallayer becomes more beneficial. In almost all cases, the Min-phys-hop algorithm manifests a highly attractive trade-offbetween the two routing policies.

5.4 Performance evaluation

So far, we have seen that the Min-phys-hop algorithm consti-tutes an appealing layer-preference policy, with low connec-tion blocking and resource consumption. However, we haverestricted ourselves to ideal networks, consisting of opticalnodes of unlimited wavelength conversion capability. In thissection, we broaden the scope of our simulations: we exam-ine a real network topology consisting of realistic opticaldevices with limited OEO conversion capability. Our simu-lations are aimed at measuring the blocking probability (theeffective measure of the goodness of the layer-preferencepolicy) and the average path length and number of physi-cal-level loops, as produced by the Min-phys-hop algorithmcompared to the entire spectrum of layer-preference policiesresiding between pushing routing completely into the lowerlayer (α = 0) and the higher layer (α = 1).

The parameters of the simulations were chosen as fol-lows: the topology we used was a real network, the 28 nodeEuropean reference network [24]; the number of wavelengthsper optical link was varied between 2 and 32 (although, dueto space limitations, we do not include all results), and thecapacity of wavelength channels was chosen as 100 units.LSP requests were generated one-by-one, while the corre-sponding source and destination nodes were selected accord-ing to a uniform distribution over all pairs of nodes. Requestsarrived according to independent Poisson processes for eachsource-destination pair and holding times were distributedexponentially with an expected value of 10 units. The aver-age request arrival intensity and the bandwidth of wavelengthchannels were selected so that there are always at least fourrequests alive between a particular source and destinationpair at the same time. The average request size was distrib-uted uniformly between 24 and 26 units.

The average ratio of blocked calls over all source-destina-tion pairs for different number of wavelengths per optical linkis depicted for unlimited wavelength conversion (M = ∞)in Fig. 10a (for 4 and 8 wavelengths per link) and 10b (for16 and 32 wavelengths per link) and for limited wavelengthand OEO conversion in Fig. 10c and 10d (for M = 1) andin Fig. 10e and 10f (for M = 2). The case when there is nowavelength conversion is left for further study. Note that therequest intensities and holding times were scaled measure-ment by measurement in order to assure that the range of callblocking rates stays sane. Therefore, it does not make senseto compare blocking ratios across simulations for differentwavelength numbers or wavelength conversion parameters.

The reason for this is that we only wanted to demonstrate that,for any choice of input parameters, the Min-phys-hop algo-rithm produces acceptable, quasi-optimal blocking ratio andresource usage and, as evidenced by the simulation results,this is exactly the case. Observe that, for a specific com-bination of wavelength number and M (OEO conversioncapability), the Min-phys-hop algorithm usually attains theblocking ratio corresponding to the best choice of the layer-preference policy (that is, the setting of α that produces theminimal blocking ratio), and this is regardless of the OEOconversion capability available in the network. For networkswith unlimited wavelength conversion, Min-phys-hop onlyapproximates the optimum, but for limited wavelength con-version, where excessively long paths are even more costlyin terms of optical transmitters and receivers, Min-phys-hopeven outperforms that. However, it is also educational toobserve that there does not seem to exist a universal one-fits-all α parameter, but instead, the best policy depends onthe actual parameters of the network.

The diagrams describing the average path length (inFig. 11a) and the number of physical-level loops (in Fig. 11b)demonstrate that not just that the Min-phys-hop algorithmproduces low blocking, but it also achieves that near a rel-atively low resource consumption when compared to otherlayer-preference policies.

6 Conclusions

In this paper, we have developed the Min-phys-hop rout-ing and wavelength assignment algorithm and examined thepractical issues concerning its deployment in GMPLS net-works. First, we introduced and analyzed layer-preferencepolicies in multi-layer networks and we identified theirrespective disadvantages, wavelength fragmentation and theemergence of physical-level routing loops. Based on theseobservations, we developed the Min-phys-hop algorithm, aheuristic layer-preference policy, and we discussed the apt-ness of the algorithm to the state-of-the-art GMPLS stan-dards, above all, the GMPLS routing extensions to OSPF-TE.We concluded that the Min-phys-hop algorithm presentsitself as a viable choice for routing and wavelength assign-ment. In order to affirm this claim, we sketched two possi-ble reference deployment scenarios. Finally, we showed thatthe Min-phys-hop algorithm is not only practical but quiteefficient too: we developed a new graph model able to cap-ture all the limitations and restrictions inherent to today’soptical switching hardware and, using this model, we con-ducted comprehensive simulation studies. The results con-firm that our algorithm reduces the number of blocked LSPrequests and uses network resources more efficiently underalmost all possible selections of input parameters, and this isregardless of the wavelength and OEO conversion capability

123

Photon Netw Commun

0

1

2

3

4

5

6

7

0.05 0.1

0.15 0.2

0.25 0.3

0.35 0.4

0.45 0.5

0.55 0.6

0.65 0.7

0.75 0.8

0.85 0.9

0.95

Alpha

Ave

rag

e p

ath

len

gth

[h

op

]

Min-HopMin-phys-hop

0

2000

4000

6000

8000

10000

12000

14000

16000

18000

20000

0.05 0.1

0.15 0.2

0.25 0.3

0.35 0.4

0.45 0.5

0.55 0.6

0.65 0.7

0.75 0.8

0.85 0.9

0.95

Alpha

Ave

rag

e n

um

ber

of

loo

ps

Min-hopMin-phys-hop

(a) (b)

Fig. 11 Average path length (a) and average number of physical level loops (b) produced by different layer-preference policies and the Min-phys-hop in a network of unlimited wavelength conversion

present in the network. We also showed that there does notexist a universally optimal static layer-preference policy andthat the Min-phys-hop algorithm realizes an adequate heu-ristics even considering the realistic limitations of contem-porary network devices.

Acknowledgements The work reported in this paper has been donewithin the framework of the European FP6 IST Project IP NOBEL II(www.ist-nobel.org).

References

[1] Banerjee, A., Drake, J., Lang, J., Turner, B., Awduche, D.,Berger, L., Kompella, K., Rekhter, Y.: Generalized multiproto-col label switching: an overview of signaling enhancements andrecovery techniques. IEEE Communications Magazine, 144–151(2001)

[2] Mannie, E., et al.: Generalized multi-protocol label switching(GMPLS) architecture, RFC 3945, October (2004)

[3] Banerjee, A., Drake, J., Lang, J., Turner, B., Kompella, K.,Rekhter, Y.: Generalized multiprotocol label switching: An over-view of routing and management enhancements. IEEE Commu-nications Magazine. 144–150, January (2001)

[4] Berger, L. (ed.): Generalized Multi-Protocol Label Switching(GMPLS) Signaling Functional Description, RFC 3471, January(2003)

[5] Kompella, K., Rekhter, Y. (eds.): Routing Extensions in Supportof Generalized Multi-Protocol Label Switching (GMPLS), RFC4202, October (2005)

[6] Kompella, K., Rekhter, Y.: OSPF Extensions in Support of Gen-eralized Multi-Protocol Label Switching, RFC 4203, October(2005)

[7] Katz, D., Kompella, K., Yeung, D.: Traffic Engineering Exten-sions to OSPF Version 2, RFC 3630, September (2003)

[8] Kompella, K., Rekhter, Y. (eds.): Intermediate System to Interme-diate System (IS-IS) Extensions in Support of Generalized Multi-Protocol Label Switching (GMPLS), RFC 4205, October (2005)

[9] Zang, H., Jue, J.P., Mukherjee, B.: A review of routing andwavelength assignment approaches for wavelength-routed opti-cal WDM networks. Opt. Netw. Mag. 1(1), 47–60 (2000)

[10] Iovanna, P., Sabella, R., Settembre, M.: A traffic engineeringsystem for multilayer networks based on the GMPLS para-digm. IEEE. Netw. 17(2), 28–37 (2003)

[11] Sabella, R., Iovanna, P., Oriolo, G., D’Aprile, P.: Strategy fordynamic routing and grooming of data flows into lightpaths innew generation network based on the GMPLS paradigm. J. Pho-tonic Netw. Commun. 7(2), 131–144 (2004)

[12] Zang, H., Jue, J.P., Sahasrabuddhe, L., Ramamurthy, R., Mukher-jee, B.: Dynamic lightpath establishment in wavelength-routedWDM networks. IEEE. Commun. Mag. 39(9), 100–108 (2001)

[13] Zhu, K., Mukherjee, B.: On-line approaches for provisioningconnections of different bandwidth granularities in WDM meshnetworks. In: Technical Digest, Optical Fiber CommunicationsConference (OFC 2002), Anaheim, CA, 549–551, March (2002)

[14] Zhu, H., Zang, H., Zhu, K., Mukherjee, B.: A novel generic graphmodel for traffic grooming in heterogeneous WDM mesh net-works. IEEE/ACM TON 11, 285–299 (2003)

[15] Zhu, H., Zang, H., Zhu, K., Mukherjee, B.: Dynamic traffic groom-ing in WDM mesh networks using a novel graph model. In: GLO-BECOM 2002 - IEEE Global Telecommunications Conference,vol. 21, pp. 2688–2692 (Nov 2002)

[16] Comellas, J., Martinez, R., Prat, J., Sales, V., Junyent, G.: Inte-grated IP/WDM Routing in GMPLS-Based Optical Networks. In:IEEE Network, vol. 17, no. 2, pp. 22–27 (2003)

[17] Retvari, G., Fodor, P., Tapolcai, J., Cinkler, T.: Challenges of multi-layer traffic engineering in GMPLS Networks. IEEE ICTON 2005,7th International Conference on Transparent Optical Networks,Barcelona, Spain, July 3–7 (2005)

[18] Garey, M., Johnson, D.: Computers and Intractability A Guideto the Theory of NP-Completeness. WH Freeman & Co, SanFrancisco, CA, USA (1979)

[19] Farrel, A. ed., Ayyangar, A., Vasseur, J.P.: Inter domain Multipro-tocol Label Switching (MPLS) and Generalized MPLS (GMPLS)Traffic Engineering - RSVP-TE extensions, RFC5151, September(2007)

[20] Farrel, A., Vasseur, J.-P., Ash, J.: A path computation element(PCE)-based architecture, RFC 4655, August (2006)

[21] Kompella, K., Rekhter, Y.: Label switched paths (LSP) hierarchywith generalized multi-protocol label switching (GMPLS) trafficengineering (TE), RFC 4206, October (2005)

[22] Shiomoto, K. (ed.): Framework for MPLS-TE to GMPLSmigration, Internet Draft, draft-ietf-ccamp-mpls-gmpls-inter-work-fmwk-03.txt, August (2007)

123

Photon Netw Commun

[23] Kodialam, M.S., Lakshman, T.V.: Integrated dynamic IP andwavelength routing in IP over WDM networks, in INFOCOM,2001, 358–366

[24] Inkret, R., et al.: Advanced Infrastructure for Photonic Networks:Extended Final Report of COST Action 266. Faculty of ElectricalEngineering and Computing University of Zagreb, (2003)

Author Biographies

Péter Fodor received his M.Sc.degree in Electrical Engineering fromthe Budapest University of Technol-ogy and Economics (BME), Buda-pest, Hungary, in 2001. He is cur-rently a consultant manager at theDepartment of Telecommunicationsand Media Informatics and a mem-ber of the High Speed Networks Lab-oratory. His research interests focuson network controlling, modeling andgeneralization.

Gábor Enyedi received anM.Sc. degree in Computer Sci-ence at the University of Tech-nology and Economics (BME) in2006. Since then, he has beena Ph.D. student at BME and amember of High Speed NetworksLaboratory at the Department ofTelecommunication and MediaInformatics. His resource inter-ests include Traffic Engineeringin circuit switched networks andIP Fast Reroute.

Gábor Rétvári received theM.Sc. and Ph.D. degrees inElectrical Engineering from theBudapest University of Tech-nology and Economics (BME),Budapest, Hungary, in 1999and 2007, respectively. He isnow a Research Assistant atthe High Speed NetworksLaboratory, Department of Tele-communications and MediaInformatics, BME. His researchinterests include QoS routing,Traffic Engineering and the net-working applications of compu-

tational geometry and the mathematical theory of network flows. He isa Perl expert, maintaining numerous open source scientific tools writtenin Perl, C and Haskell.

Tibor Cinkler [M’95] hasreceived M.Sc. (94) and Ph.D.(’99) degrees from the Buda-pest University of Technologyand Economics (BME), Hun-gary, where he is currentlyassociate professor at the Depart-ment of Telecommunicationsand Media Informatics. Hisresearch interests focus on opti-misation of routing, traffic engi-neering, design, configuration,dimensioning and resilience ofIP, Ethernet, MPLS, ngSDH,OTN and particularly of hetero-

geneous GMPLScontrolled WDM-based multilayer networks. He isauthor of over 130 refereed scientific publications and of three patents.He has been involved in numerous related European and Hungarianprojects including ACTS METON and DEMON; COST 266, 291, 293;IP NOBEL I and II and MUSE;NoE e-Photon/ONe andNoE e-Pho-ton/ONe+;CELTICPROMISE; NKFP, GVOP, ETIK; and he ismemberof ONDM, DRCN, Broad-Nets, AccessNets, IEEE ICC and Globecom,EUNICE, CHINACOM, Networks,WynSys, ICTON, etc. Scientific andProgramm Committees. He has been guest editor of a Feature Topic ofthe IEEE ComMag and reviewer for many journals.

123