Experimental evaluation of a VON Resource Brokerand Compositor for deployment and performance of

virtual GMPLS control planesRicard Vilalta, Raul Munoz, Ramon Casellas and Ricardo Martinez

Centre Tecnologic de Telecomunicacions de Catalunya (CTTC)Parc Mediterrani de la Tecnologia, Av. Carl Friedrich Gauss 7

08860 Castelldefels (Barcelona), Spain.Email: {rvilalta, rmunoz, rcasellas, rmartinez}@cttc.es

Abstract—We present a Virtual Optical Network (VON) Re-source Broker and Compositor architecture to dynamicallyconfigure and deploy independent instances of virtual GMPLS-controlled WSON. The VON Resource Broker consists of fourblocks: resource manager, VON controller, resource allocator andresource configurator. In this paper, we evaluate the impact of theproposed VON Resource Broker on the provisioning of virtualGMPLS control planes and the performance of provisioninglightpaths within a virtual control plane through experimentationin the ADRENALINE Testbed. The obtained results show thetrade off between the diminution of blocking rate of VONrequests, through the provision of more virtual resources forthe virtual control planes, and the deterioration of the virtualcontrol plane performance (e.g. higher setup delay and higherblocking probability while provisioning lightpaths).

I. INTRODUCTION

Virtual Optical Networks (VON) support dynamic provi-sioning of dedicated networks over the same network infras-tructure, which has received a lot of attention by networkinfrastructure providers. The stringent network requirementsof the emerging high bandwidth and dynamic applicationssuch as high-definition video streaming (e.g., telepresence,television, remote surgery, etc.), and cloud computing (e.g.,real-time data backup, remote desktop, etc.) can be supportedby the deployment of dynamic infrastructure services to buildad-hoc VON services, which is known as infrastructure as aservice (IaaS) [1]. Thus, network service providers can request,on a per-need basis, a dedicated VON for each application andhave full control over it (Fig. 1).

Optical network virtualization technologies allow the parti-tioning/composition of the network infrastructure (i.e., physi-cal optical nodes and links) into independent virtual resources,adopting the same functionality as the physical resource [2].The composition of these virtual resources (i.e. virtual opticalnodes and links) allows the deployment of multiple VirtualOptical Networks (VON). A VON must be composed of notonly a virtual transport plane but also of a virtual control plane,with the purpose of providing the required independent andfull control functionalities (i.e., optical connection provision-ing, traffic engineering, protection/restauration, etc.).

A virtual GMPLS control plane is a distributed entity

composed of Virtual Connection Controllers (VCC) (one pervirtual optical switch) executing several collaborative pro-cesses and a Data Communication Network (DCN) based onvirtual IP Control Channels (IPCC) to allow the exchangeof control messages between the VCCs (see Fig. 2). Theprocesses running in a VCC are RSVP-TE signalling agentfor connection provisioning, OSPF-TE routing agent for topol-ogy and resource dissemination, the Link Resource Manager(LRM) agent for local resource management and HardwareAbstraction Layer (HAL) agent for managing the ConnectionController Interface (CCI).

This paper presents and evaluates a VON Resource Bro-ker and Compositor for dynamic GMPLS-controlled WSONinfrastructure services, whose task is to dynamically de-ploy GMPLS-controlled VONs from service provider requestsand serves as interface between service and infrastructureproviders. This VON Resource Broker and Compositor wasfirst presented by the authors in [3]. This paper also evaluatesthe effects that the introduction of the virtual resources by theResource Broker and Compositor may imply to the deployedVirtual Optical Networks (i.e. higher setup delays and higherblocking probabilities).

The remainder of the paper is organized as follows. InSection 2, we give an overview of the state of the art ofResource Brokers. Section 3 describes the proposed VONResource Broker and Compositor architecture and how virtualdata and control plane resources are managed. Section 4describes the results obtained from applying the proposedVON Resource Broker and Compositor architecture to theADRENALINE Testbed and its impact on virtual control planedeployment and performance. Finally, Section 5 concludes.

II. BACKGROUND

A Virtual Optical Network is the composition of isolatedvirtual optical resources, which are existing simultaneouslyover the same physical optical network infrastructure [4]. Inthis section, an analysis of the several methods for partitioningthe physical optical network and composition the isolatedvirtual optical networks is provided. These methods for par-titioning and compositing the VONs are generally refereed

Virtual Control Plane  Layer 

VON1 VON2

Virtual IP Control Channel

Virtual Transport Plane Layer

Virtual GMPLS 

Controller

Control Channel

Virtual Transport Plane Layer

Virtual Optical Switch

Virtual Optical Link

Physical GMPLS‐controlled

Optical Network Infrastructure 

GMPLS controlled optical Node

Fig. 1. VON layer architecture

Virtual Nodes

Virtualization Server

ROADM/OXC

Physical Node

IPCC IPCC

Virtual ROADM/OXC

Virtualization Layer

Virtual GMPLS

Controller #1

Virtual GMPLS

Controller #N

LD

Link Resource Manager

Connection Controller GMPLS RSVP-TE

TED

GMPLS OSPF-TE

Path Computation

HAL

Routing Controller

1G PCIe RJ45

Vir

tual

GM

PLS

Co

ntr

olle

r

…

IP FORWARDING

1G PCIe RJ45

FIB

LD Link Resource

Manager

1G PCIe RJ45

Vir

uta

l IP

CC

Vit

ual

IPC

C

CC

I

CCI

…

Virtual GMPLS Controller #1

Virtual GMPLS Controller #2

Virtual GMPLS Controller #N

Fig. 2. Partitioning of physical transport and control resources

as Resource Brokers. In the first subsection, an introductionto VON Resource Broker taxonomy is discussed, and inthe second subsection several literature resource brokers areanalyzed.

A. VON Resource Broker Taxonomy

This subsection describes the different properties that Re-source Brokers have when providing the solution to theproblem of deploying virtual optical networks on top of acommon shared optical infrastructure.

One of the most important properties for a Resource Brokeris the form it processes incoming requests, either in a static ordynamic way. Static request processing implies pre-planningof the resources, while dynamic request processing consists onthe resource allocation depending on the state of the availableoptical infrastructure.

Another important property for the Resource Broker is theownership of the optical resources, i.e. exclusive or shared.Exclusive optical resources shall be assigned to a singleuser/request, while shared resources can be allocated to severalusers/requests. Examples of possible shared optical resourcesare optical nodes and optical fibers, while wavelength channels

within a link shall be considered exclusive.Several interfaces to the Resource Broker can be observed,

being the most common an interface for requesting virtualoptical networks based on XML requests. These interfaceshave not been standardized and a possible general solutionwould be the usage of Network Description Language, whichhas been successfully applied to optical networks [5].

The last important characteristic of the Resource Broker isthe allocation of resources to deploy a control plane for each ofthe different virtual optical networks. The requests for VONsshall include key control plane configuration parameters.

B. Studied Resource Brokers

The authors in [6] propose a VON service compositionframework which focuses on shared VON provisioning acrossmultiple domains for Grid applications. The presented resourcemanagement focuses on shared VON resources allowing dy-namic explicit lightpath provisioning by user. We have ana-lyzed the solution proposed for a single domain VON service,which consists of a layer acting as middleware, for whichthe Service Access Interface (SAI) is introduced. The SAIinteracts with the Physical Optical Network Traffic EngineryDatabase (PON-TED) and the VON-TED. This single domainservice is responsible to interact with the Control Plane of theInfrastructure Provider.

[7] proposes an impairment-aware Virtual Optical NetworkComposition Mechanism, which is composed of an opticalphysical infrastructure layer (i.e. optical network resources),an optical network composition layer (i.e. where the physicalresources are first abstracted and then partitioned/aggregatedinto virtual resources) and an optical network control andmanagement layer. The focus of this paper is on the resourceallocation of the data plane optical resources taking intoaccount Physical Layer Impairments (PLIs).

In the studied Resource Brokers, the virtualization of Op-tical Networks only includes the virtualization of the opticaldata plane, not taking into account the necessary control planefor each deployed virtual optical network. Neither of the pre-sented VON Resource Broker architectures take into accountthe necessity of a control plane for the deployed VON. Ourproposed GMPLS-controlled virtual optical network resourcebroker and compositor takes into account the necessity of theprovisioning of a virtual Control Plane, responsible of theVON, while its is also able to provision the virtual transportplane.

III. VIRTUAL OPTICAL NETWORK BROKER ANDCOMPOSITOR

The proposed GMPLS-controlled VON resource broker andcompositor architecture (Fig. 3), consists of the followingfour blocks: the Resource Manager, VON Controller, ResourceAllocator and Resource Configurator.

The Resource Manager handles both the virtual control andtransport resources. As for the virtual control resources, itmanages the available IP subnetworks that shall be used toestablish the virtual IPCC to later deploy dedicated DCN

All-Optical Wavelength

Switched Optical Network

VON Resource Broker and Service Compositor

Virtualization Servers

VON Controller

Resource Allocator

Resource Manager

Resource Configurator

Scenario Configuration

File

ADNETCONF

VON #1

VON #N

…

VON Request

Fig. 3. GMPLS-controlled VON Broker and Compositor Architecture

for each virtual control plane. It is also responsible formanaging the number and location of the available virtualGMPLS controllers that each physical GMPLS controller (i.e.,virtualization servers in the ADRENALINE testbed) supports(static partitioning), as well as their configuration information(including the management IP address, the amount of CPUpower and available RAM). The Resource Manager also storesall the information required to configure the processes runningin the virtual GMPLS controllers. For example, it stores theIP addresses and users (login and passwords) of the opticalswitches (each switch has a maximum number of users thatequals to the maximum number of partitions supported), theset of available GMPLS router addresses and identifiers, thepath computation algorithm identifiers, the set of wavelengthassignment algorithm identifiers or the default protocol timers(e.g., hello, refresh) or the set of Path Computation Element(PCE) IP addresses to request path computation. Transportplane resources are described as optical links and switches.Each optical switch has a determined number of input andoutput ports and a maximum number of partitions is feasible.The specific partitioning of an optical switch is performeddynamically, based on the requirements of the requested VON.Each optical link stores the information about the edge opticalswitches the link is connected to, the number of supportedwavelengths and their identifiers, as well as any link parameter(i.e. optical impairments). The partitioning of an optical linkis on a wavelength granularity basis.

The VON Controller accepts incoming TCP sessions, whichinclude VON requests, and handles these requests asyn-chronously and dynamically. Once the VON identifier, whichis a unique identifier per each VON, is assigned, or found inthe request database, the VON controller triggers the resourceallocator in order to process the VON request, which consistson the allocation of the new VON (i.e., VON deployment),the modification of resources assigned to an existing VON orthe releasing of the resources in case a VON is torn down.A VON request is modeled as a graph that describes a set

Fig. 4. VON A and VON B in ADRENALINE Testbed

of virtual optical switches and links for the virtual transportplane, specifying for each one the number of requested inputand output optical ports, and the number of wavelengths re-spectively. The VON request also includes some requirementsfor the virtual control plane, such as the needed capacities(i.e. CPU power, RAM) for the virtual GMPLS controllers,or the configuration values for the control processes of thevirtual GMPLS controllers, which can be later modified bythe service provider.

The resource allocator assigns the virtual transport andcontrol resources to the requested VON. For the virtual controlplane, it allocates the virtual GMPLS controllers, based onthe requested CPU power and RAM, and assigns the GMPLSrouter address. It also assigns IP addresses and GRE tunnelsfor the required IP Control Channels (it is assumed thatan IPCCs are assigned following the same topology as thetransport plane). The configuration values for all GMPLSprocesses running in a virtual GMPLS controller are alsoassigned. For example, for the HAL process, it is specifiedthe set of allocated wavelengths per link. The configuration forthe LRM process is generated, including for each TE link theswitching capabilities and the maximum and available band-width, depending on the allocated wavelengths. The OSPF-TE process is configured with the adjacent router addressesand the interfaces that shall be used for the VON topologyand resource dissemination. For the virtual transport plane,the target is to assign the requested number of availablewavelengths (N) for each virtual optical link of the requestedVON. Several wavelength assignment schemes can be applied,such as First-Fit (select the first N wavelengths being availableand common in all requested virtual optical links), or takinginto account physical optical impairments [7].

The resource configurator generates the virtual transportand control plane configuration XML file, which describesa VON scenario model that can be set up, modified or torndown by means of ADRENALINE Network Configurator(ADNETCONF) [8] (see Fig. 4), which is a software toolin charge of scenario model management in ADRENALINETestbed. With ADNETCONF, the scenario model is thenserialized to the formal representation of the scenario thatthe processing engine understands. Up to five different XML

OXC

1

R-OADM 2

SMF G.652 (35 Km.)

32 λ DWDM

NZDSF+ G.655

(150 Km.)

32 λ DWDM OXC

2

SMF G.652 (35 Km.)

32 λ DWDM

SMF G.652 (35 Km.)

32 λ DWDM

SMF G.652 (50 Km.)

32 λ DWDM

R-OADM 1

ADENALINE Testbed

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

VON B

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

VON A

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

VON D

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

VON C

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

VON F

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

4 λ DWDM

VON E

Fig. 5. ADRENALINE Testbed WSON scenario and multiple VONs

files are produced; one describing the logical DCN topologyfor the virtual control plane, and the others describing theconfiguration of the different GMPLS processes.

IV. EXPERIMENTAL PERFORMANCE EVALUATION

The experimental evaluation consists of evaluating the im-pact of the allocation of virtual control resources (i.e. GMPLSControllers) when the VON Resource Broker and Compositoris requested to deploy different VONs on the ADRENALINETestbed.

The ADRENALINE Testbed [9] is a GMPLS-based Intelli-gent Optical Network composed of an all-optical WSON with2 ROADMs and 2 OXCs providing reconfigurable (in spaceand in frequency) end-to-end lightpaths, deploying a total of610 km of G.652 and G.655 optical fiber (see Fig. 5). Eachoptical node is also equipped with a virtualization server (i.e.,physical GMPLS controller) running in a Linux-based routerwith an Intel Core 2 Duo E6550 2.33 GHz processor and 1024MB of RAM. Each node virtualization server, is running upto 6 KVM Virtual Machines (VMs), that can be started orstopped by the VON Resource Broker and Compositor.

These virtual machines have the same amount of RAM(which is the RAM of the virtualization server divided bythe number of active virtual machines) and share the CPUpower of the server. The virtualization of the GMPLS controlplane of the ADRENALINE Testbed to partition the physicalcontrol plane into multiple virtual control plane instances hasbeen previously addressed by the authors in [10].

The performed evaluation have only taken into account theperformance measurements for the control plane, so that therehas been no deployment of the configurations for the transportplane. We have considered 32 wavelengths per each opticallink.

A. Impact of VON Resource Broker and Compositor on virtualcontrol plane deployment

In order to evaluate the performance of VON provision-ing, the performance values are obtained when dynamically

Fig. 6. Blocking rate of VON requests

Fig. 7. Setup and tear down mean times for VON in the ADRENALINETestbed

provisioning VONs, in which the inter-arrival (IAT) processis Poisson, and the holding time (HT) follows a negativeexponential distribution. The mean inter-arrival time is set to3 seconds and the average VON-holding time is varied duringall different experimentations, yielding an offered traffic loadfrom 0.1 to 50 Erlangs. The topology for each VON request israndomly selected from the space of feasible topologies, whichis determined by the physical optical network infrastructure.For the ADRENALINE Testbed, we consider a set of 28topologies.

We consider three different scenarios for the evaluationof the VON Resource Broker and Compositor performancewhen deploying VONs. In the first scenario, each virtualizationserver has 2 virtual machines running on top of it, whilst inthe second scenario 4 virtual machines are running on top ofthe virtualization server and in the third scenario, 6 virtualmachines.

The number of wavelengths requested for each VON isset to 4, and the wavelength assignment algorithm used isFirst-Fit. 104 VONs have been requested for measuring theblocking rate, which has been computed in the 3 presentedscenarios. In Figure 6, we can observe how the blocking rateof VON requests is lower when we are using more virtualmachines per virtualization server. For a given VON requestload of 1 Erlangs, the obtained blocking rate of VON requestsis 0.02% in the case of using 6 virtual machines and 19.15%in the case of using only 2 virtual machines. These results arecoherent with the expectation of lesser blocking probability incase of providing more virtual machines, which act as GMPLS

Fig. 8. Blocking probability in VON A

Fig. 9. Setup Delay for connection requests in VON A

controllers.We also want to evaluate the impact of the VON Resource

Broker and Compositor on the setup and tear down delay whendeploying the VON requests, only measuring the setup and teardown delay of the control plane, depending on the numberof virtual machines in the virtualization servers. 100 requestshave been performed for each data point. Figure 7 shows howthe necessary time for setting up a VON is increased linearlywhen the number of virtual machines in the virtualizationservers increases (e.g. setup time is 20.97 seconds with 2 VMs,whilst it is 22.10 seconds for 4 VMs). Instead, we observe thatthe tear down time does not increase with regard the numberof virtual machines. These results can be explained becausefor setting up a VON and its control plane a higher amount ofCPU and RAM resources need to be used, whilst for tearingdown the VON this amount is not such significant.

B. Impact of allocated virtual resources on virtual controlplane performance

In order to evaluate the impact of the allocated virtual re-sources on virtual control plane performance we have consid-ered 4 scenarios, depending on the number of VON deployedon the ADRENALINE Testbed. Figure 5 depicts the scenariowhere 6 VONs have been deployed on the ADRENALINETestbed. Scenarios for 4, 2 and 1 VON deployed have alsobeen considered. Each VON has been assigned with 4 light-paths for each of its virtual optical links.

The performance values are obtained when dynamically pro-visioning connections inside each VON, the lightpath requestinter-arrival (IAT) process is Poisson, and the holding time

(HT) follows a negative exponential distribution. The meaninter-arrival time is set to 3 seconds and the average VON-holding time is varied during all different experimentations,yielding an offered traffic load from 2 to 50 Erlangs. 104

connections have been requested inside each VON for the 4scenarios.

In Figure 8, we can observe how the blocking probabilityinside the VON A (for the other active VONs the results areequivalent) is significantly affected depending on the numberof deployed VON virtual control planes, which depends on thenumber of VM in the virtualization servers upon each opticalnode. For example the Blocking Probability (BP) in VON Ais 0.72% when the traffic load is of 5 Erlangs and only VONA is active, whilst the BP in VON A for the same traffic loadis 1.74% when 6 VONs are deployed on the ADRENALINETestbed. The obtained results can be explained because of theincrease of the Setup Delay (SD), implies a longer lighpathestablishment time, so there are more concurrent requestsbeing processed at the time.

Figure 9 shows the mean Setup Delay in VON A and howit is affected depending the number of deployed VON virtualcontrol planes. We observe, that the SD in VON A is oneorder of magnitude higher (337 ms) in the case of 6 VONcontrol planes than in the case of 2 VON control planes (32.61ms) for a traffic load in both VONs of 5 Erlangs. This poorperformance is explained by the virtualization of the GMPLScontrollers and the amount of GMPLS controllers running ontop of a virtualization server. The obtained results of SD inthe case of 6 VON virtual control planes (337 ms / 5 Erlangs)might be suitable for lightpath provisioning, but they are nottolerable in case of connection restauration.

V. CONCLUSIONS

We have presented the VON Resource Broker and Com-positor which acts as interface between service providers andinfrastructure providers to deploy virtual GMPLS-controlledWSON infrastructure services over the same physical opticalnetwork infrastructure. Experimental evaluation carried outin the ADRENALINE Testbed has shown the feasibility ofdeploying independent instances of GMPLS control plane foreach deployed virtual transport network under dynamic VONrequests, providing low delays for VON setup and tear down.

The measurement of the impact of deploying several virtualVON GMPLS control planes on its performance (i.e. setupdelay and blocking probability) has resulted in higher setupdelay and higher blocking probability in lightpath provision-ing. These obtained results show the trade off between thediminution of blocking rate of VON requests, through theprovision of more virtual resources for the virtual controlplanes, and the deterioration of the virtual control planeperformance.

In case we want to obtain a maximum setup delay of 150msand a blocking probability lower than 2% while requestinglightpaths to a VON with 5 Erlangs traffic load, the maximumnumber of virtual machines to be deployed in a virtualizationserver are 4 VMs, which will result in a blocking rate of

1.21% for a VON request rate of 1 Erlang in the presentedADRENALINE Testbed.

ACKNOWLEDGMENT

This work was partially funded by the MICINN (SpanishMinistry of Science and Innovation) through the project DO-RADO (TEC2009-07995).

REFERENCES

[1] E. Keller and J. Rexford, “The ”platform as a service” model fornetworking,” in Proceedings of the 2010 internet network managementconference on Research on enterprise networking, ser. INM/WREN’10,2010.

[2] R. Nejabati et al., “Optical network virtualization,” in ONDM 2011.[3] R. Vilalta et al., “Experimental demonstration of a virtual optical

network resource broker and compositor for dynamic gmpls wsoninfrastructure services,” in OFC/NFOEC, 2012.

[4] R. Nejabati, E. Escalona, S. Peng, and D. Simeonidou, “Optical networkvirtualization,” in Optical Network Design and Modeling (ONDM), 201115th International Conference on, feb. 2011, pp. 1 –5.

[5] J. van der Ham et al., “Using the network description language inoptical networks,” in Integrated Network Management, 2007. IM ’07.10th IFIP/IEEE International Symposium on, 2007, pp. 199 –205.

[6] Y. Wang, Y. Jin, W. Guo, W. Sun, and W. Hu, “Virtualized opticalnetwork services across multiple domains for grid applications,” Com-munications Magazine, IEEE, vol. 49, no. 5, pp. 92 –101, may 2011.

[7] S. Peng, R. Nejabati, S. Azodolmolky, E. Escalona, and D. Simeonidou,“An impairment-aware virtual optical network composition mechanismfor future internet,” in European Conference and Exposition on OpticalCommunications, 2011.

[8] F. Galan Marquez and R. Munoz, “An automatic model-based recon-figuration and monitoring mechanism for flexible gmpls-based opticalnetworking testbeds,” in Optical Network Design and Modeling, ser.Lecture Notes in Computer Science, I. Tomkos, F. Neri, J. Sole Pareta,X. Masip Bruin, and S. Sanchez Lopez, Eds. Springer Berlin /Heidelberg, 2007, vol. 4534, pp. 239–248.

[9] R. Munoz, C. Pinart, R. Martinez, J. Sorribes, G. Junyent, M. Maier, andA. Amrani, “The ADRENALINE Test Bed: Integrating GMPLS, XMLand SNMP in transparent DWDM networks,” in IEEE CommunicationsMagazine, vol. 43, no. 8, August 2005, pp. 40–48.

[10] R. Munoz, R. Casellas, R. Martınez, R. Vilalta, J. Vılchez, andJ. Vazquez, “Virtualizing adrenaline testbed for deploying dynamicgmpls-controlled wson as a service,” in OFC/NFOEC, 2011.