A258 J. OPT. COMMUN. NETW./VOL. 1, NO. 2 /JULY 2009 Agraz et al.

Au

Design and Implementation of aGMPLS-Controlled Grooming-Capable

Optical Transport NetworkFernando Agraz, Luis Velasco, Jordi Perelló, Marc Ruiz, Salvatore Spadaro,

Gabriel Junyent, and Jaume Comellas

ssIwbtap

(bisfbtsasmpAww

gmthrdnegv1iwb

t

Abstract—Aiming at better resource utilization, animportant requirement of future optical transportnetworks is the capability to accommodate subwave-length client flows efficiently. This can be put into ac-tion thanks to the enhanced traffic engineering (TE)protocols provided within the generalized multipro-tocol label switching (GMPLS) standardization. Thepresent paper concentrates on the design and imple-mentation of a GMPLS-controlled grooming-capabletransport infrastructure, namely, the automaticallyswitched optical network (ASON)ÕGMPLS CARISMAtest bed. Through the paper, the operation of aGMPLS-controlled multilayer network architectureis introduced, subsequently highlighting implementa-tion issues that come to light. Special attention is de-voted to a centralized flow reallocation module de-ployed in the CARISMA test bed to minimize theoverall network cost. In this context, an integer lin-ear programming (ILP) formulation to obtain its op-timal cost is derived and low-weighted metaheuris-tics providing a nearly optimal solution areadditionally proposed. All contributions in the paperare supported by illustrative experimental results.

Index Terms—Assignment and routing algorithms;Networks, circuit-switched; Network optimization.

I. INTRODUCTION

W avelength-routed optical networks have re-ceived increasing attention as a promising ap-

proach to deploy end-to-end transparent networks in acost-effective way. Their main goal is to optically by-pass highly overloaded electronic routers, which re-sults in a significant reduction in optical-to-electrical(O/E) ports, thus decreasing the overall network cost.Such networks, however, have traditionally beenrather static, due to the manual optical circuit provi-

Manuscript received November 7, 2008; revised January 23, 2008;accepted February 18, 2009; published July 1, 2009 �Doc. ID 103862�.

The authors are with the Advanced Broadband CommunicationsCenter (CCABA), Universitat Politècnica de Catalunya (UPC), JordiGirona 1-3, 08034 Barcelona, Spain (e-mail: perello@ac.upc.edu).

Digital Object Identifier 10.1364/JOCN.1.00A258

1943-0620/09/02A258-12/$15.00 ©

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

ioning process. With the advent of the automaticallywitched optical network (ASON) architecture [1], theTU-T has enhanced wavelength-routed optical net-orks with dynamic connection capability. This capa-ility is accomplished by means of a control plane en-ity, responsible for the establishment, maintenance,nd release of connections over the optical transportlane.

In parallel, the Internet Engineering Task ForceIETF) has standardized generalized multiprotocol la-el switching (GMPLS) [2] as a set of protocols tomplement a common control plane, able to manageeveral switching regions in an integrated way. Inact, not only can packet-switched-capable interfacese managed by the different GMPLS protocols, buthey can also manage time-division multiplexing [e.g,ynchronous optical network/synchronous digital hier-chy (SONET/SDH)], lambda, and even fiber-witched-capable interfaces. This makes GMPLS theost accepted solution for implementing the control

lane functionalities in the ASON architecture. TheseSON networks with a GMPLS-capable control planeill be hereafter referred to as ASON/GMPLS net-orks.

The role of IP as a convergent technology has trig-ered the development of a wide range of IP-basedultimedia services, like HDTV, video conferencing,

elemedicine applications, or Internet telephony, eachaving different bandwidth or quality of service (QoS)equirements. This huge, heterogeneous, and pre-ominantly bursty generated amount of traffic posesew challenges to network operators to provide a cost-ffective data transmission. Because the bandwidthranularity of wavelength-routed optical networks isery coarse, typically a whole wavelength supporting0 or even 40 Gbps Ethernet or SONET/SDH tributar-es, these networks lack the flexibility to support sub-avelength traffic demands, which leads to poorandwidth usage.

In this context, the term traffic grooming identifieshe process of packing several low-speed traffic

2009 Optical Society of America

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

sstpittAFsiodf

tswTcr�cfittoantsdoir�dsmprv

SnswIgIcFSmac

Agraz et al. VOL. 1, NO. 2 /JULY 2009/J. OPT. COMMUN. NETW. A259

Au

streams into higher-speed streams (e.g., see [3–5]).This problem, typically focused on reducing the num-ber of add-and-drop multiplexers (ADM) in legacySONET/SDH ring architectures [3], has also been ex-tended to maximize optical channel bandwidth usagein general WDM meshed transport networks [4,5].

From the GMPLS point of view, the grooming prob-lem is translated into merging several higher-orderlabel-switched paths (LSPs) into a lower-order LSP(e.g., grooming packet LSPs carrying IP traffic into a�-LSP). Such an LSP aggregation in GMPLS is accom-plished by advertising newly created lower-orderLSPs as forwarding adjacency LSPs (FA-LSPs, [6,7]),for instance, by means of the open shortest path first-traffic engineering (OSPF-TE) protocol [8]. In thisway, conventional data links along with the previouslyadvertised FA-LSPs can indistinctly enter the pathcomputation process. Supposing that a valid routewould be found, resource reservation would then beperformed by resource reservation protocol traffic en-gineering (RSVP-TE) [9].

During this operation, a dynamic virtual topology iscreated and modified the whole time. This virtual to-pology is comprised of those existent single- ormultiple-hop �-LSPs, along with unallocated datalinks spanning one single hop in the physical topology.Because new �-LSPs are dynamically allocated in thenetwork, and connection holding times are typicallyrandom, a suboptimal allocation of resources may ex-ist at any time. In this context, a centralized resourcereallocation module could be deployed in the network.As its main objective, the module would be responsiblefor periodically checking the occupancy of the existent�-LSPs. Thus, client connections supported on �-LSPshaving a low occupancy may be rearranged, if pos-sible, onto alternative medium-loaded �-LSPs, whichwould result in better bandwidth usage in the net-work, as well as in a release of O/E port pairs. In thisway, network resource utilization could be improvedin the network. Note that along the reallocation pro-cess, some working �-LSPs need to be rerouted. Inthis regard, [10] specifies a rerouting process called“make-before-break.” This procedure consists of estab-lishing a new LSP and transferring traffic from theold LSP to the new one before the old LSP tunnel isfinally torn down. In this way, reallocation proceduresdo not cause any working traffic disruption, which be-comes of critical importance for end users.

The goal of this paper is to introduce the design andfurther implementation of an experimental GMPLS-based grooming-capable network test bed. To this end,we first present a generic GMPLS-controlledmultilayer network architecture. With this objective,the current standardization framework is reviewed,accompanied by some illustrative examples. The nextstep comprises the design of a centralized off-line re-

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

ource reallocation module. To address this, we con-truct an integer linear programming (ILP) formula-ion to obtain the optimal results. Moreover, werovide light-weight metaheuristics for subsequentmplementation, which obtains a nearly optimal solu-ion. In the experimental evaluation, a discussion onhe implementation of the FA-LSP functionality in theSON/GMPLS CARISMA test bed is reported. Later,A-LSP performance is assessed by experimental re-ults. Furthermore, on the basis of an already operat-ng multilayer network infrastructure, a centralizedff-line resource reallocation module is additionallyeployed, experimentally quantifying its benefits inront of simple FA-LSP operation.

The reconfiguration problem in traffic-grooming op-ical networks has been covered in the literature (e.g.,ee [11–13]). The authors in [11,12] redesign the net-ork topology as soon as the offered traffic changes.wo costs are taken into consideration, the O/E portost and the reconfiguration cost. In this scheme, theesult of a reconfiguration is basically the set of-LSPs to be established, as well as the route for eachlient LSP in the new topology. Because during recon-guration some �-LSPs are set up while others areorn down, the reconfiguration cost minimizes the dis-ance between the current topology and the targetedne. Alternatively, the authors of [13] propose an ad-ptation mechanism to follow traffic variations in theetwork. Therein, a mechanism monitors the load ofhe links and sets up or tears down �-LSPs when theupported load is either higher or lower than two pre-efined thresholds. Specifically, this algorithm is peri-dically run several times per hour. In contrast to ex-stent work in the literature, our approach aims toeduce the needed O/E port pairs. Hence, no new-LSPs are created. Conversely, whether a �-LSP isecided to be torn down, the client LSPs are reroutedo that the number of allocated network resources isinimized. We observe in the experimental results

reseneted later that running the proposed metaheu-istic algorithm only a few times a day obtains highlyaluable cost reduction in the network.

The remainder of this paper continues as follows.ection II presents a GMPLS-controlled multilayeretwork architecture. Section III introduces the de-ign of a centralized resource reallocation module,hich will be afterwards implemented in the CAR-

SMA test bed. Section IV is devoted to GMPLS-basedrooming implementation and validation. SubsectionV.A presents the ASON/GMPLS CARISMA test bedharacteristics. Subsections IV.B and IV.C describeA-LSP implementation and assess their benefits.ubsection IV.D quantifies performance improve-ents that stem from using a centralized resource re-

llocation module in the network. Finally, Section Voncludes the paper.

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

stscc

bsEoofbrlwiwAtra

dnGeetcm

potapoatwptlsCt

pLttaat

A260 J. OPT. COMMUN. NETW./VOL. 1, NO. 2 /JULY 2009 Agraz et al.

Au

II. GMPLS-CONTROLLED MULTILAYER NETWORKARCHITECTURE

A. Principle of Operation

The ITU-T establishes in G.805 [14] a reference lay-ered transport network architecture with technology-independent relationships among functional entities.Therein, each network layer has a twofold role,namely, a server role to the client layer above it aswell as a client role to the network layer below it. Inbrief, a subnetwork describes the capacity to associatea set of connection points (CPs) to convey so-calledcharacteristic information. With such an objective,two possible kinds of connection are defined. A linkconnection is a fixed and inflexible connection betweentwo CPs. Conversely, a subnetwork connection (SNC)is a flexible connection that may be set up and re-leased by either the control or the management plane.As a result, a network connection is a concatenation ofsubnetwork and link connections delimited by a ter-mination connection point (TCP) pair. For the sake ofgenerality, we present FA-LSP operation in a G.805-compliant context. Recall that correspondence be-tween ITU-T and IETF terminology can be found in[15,16].

A single-layered four-node all-optical network is ex-emplified in Fig. 1(a). In such a scenario, link connec-tions (representing the different wavelength channeldata links) associate CPs at remote neighboringnodes. These link connection sets are bundled intonetwork connections between remote TCPs, which re-spectively represent dense WDM (DWDM) traffic en-gineering (TE) links and optical network ports. Let ussuppose now that a �-LSP is set up between ingressnode E and egress node D [Fig. 1(b)]. The incoming cli-ent signal at the optical node E is adapted and cross-connected by an SNC to an outgoing CP. This CP is, inturn, connected through a data link to an incoming CPin the neighbor. At the intermediate nodes F and G,an SNC binds incoming and outgoing CPs, which

Fig. 1. Example of a single-layered network architecture.

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

hould be mapped to the same wavelength in the casehat no wavelength converter is used. As soon as theignal reaches the destination node D, this one isross-connected, adapted, and sent to the optical ac-ess port.

Note that the whole process explained above woulde transparent to a hypothetical client network. Inuch an overlayed network scenario [2], nodes D and

would appear to be directly connected. As a matterf fact, each layer runs its own control plane, so thatperations in one layer become totally independentrom those in other layers. While this approach haseen typically deployed due to information exchangeestrictions across different network domains, it ofteneads to suboptimal resource allocation in the net-ork. Contrariwise, in a peer network model [2] there

s a common control plane that, having complete net-ork knowledge, governs all layers in a unified way.lternatively, an augmented model [2] would lie be-

ween the overlay and peer models, where each layeruns its own control plane instance and only a limitedmount of information is exchanged among them.

In this context, the enhanced TE protocols intro-uced in GMPLS pave the way to a peer multilayeretwork architecture, controlled by means of aMPLS-enabled common control plane. The enablingntity to this goal is the FA. In GMPLS, those already-stablished lower-layer LSPs (e.g., �-LSPs) are adver-ised as FA-LSPs, which can be used to transport newlient LSPs. In this way, lower-layer resources can beore effectively utilized.

Without loss of generality, a two-layered networkeer architecture is assumed in this work, that is, anptical server layer and a client aggregation layer onop (e.g., SONET/SDH, MPLS, GbE, etc.). This layerllows the mapping of the client traffic to be trans-orted over the DWDM physical layer. At the bottom,ptical nodes provide network ports as well as clientccess ports, used to inject an aggregated client flowo the network. The incoming signal would be after-ards adapted, switched to a network port, multi-lexed into a DWDM bundle, and finally transmittedo the next optical node. On top, the client aggregationayer includes generic nodes providing electricalwitching, flow aggregation, and many other features.lient nodes are connected to optical nodes through

he client access ports.

Figure 2 shows the architecture of the two-layeredeer network under consideration. In Fig. 2(a) a clientSP is set up between nodes E and D. Imagine that

he client LSP requested bandwidth becomes 1/4 ofhe total wavelength capacity. The incoming signal isdapted and further inserted into an outgoing �-LSPggregated signal, reaching in this way the destina-ion node D. Note that no processing is needed at cli-

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

whNlbafb

throsttt

inhfnfcrttlOtp

Agraz et al. VOL. 1, NO. 2 /JULY 2009/J. OPT. COMMUN. NETW. A261

Au

ent intermediate nodes F and G, as the signal opti-cally bypasses them through the �-LSP E–D. At thedestination, the signal is demultiplexed and the clientsignal is cross-connected, adapted, and delivered tothe sink TCP (recall, termination connection point). Inthe resulting scenario, nodes E and D appear to be di-rectly connected with an additional capacity of 3/4 ofthe total wavelength capacity. Further looking at Fig.2(b), the previously established FA-LSP has now beenused to create an additional client LSP between nodesA and G. To this end, two additional FA-LSPs betweennodes A–E and D–G are set up, providing the requiredconnectivity at the client layer. Specifically, an SNCassociates incoming CPs with outgoing CPs at inter-mediate client nodes E and D. It is worth mentioningthat FA-LSPs supported on different wavelengthchannels can be concatenated without wavelengthconverter requirements, as the signal is O/E convertedat the client layer.

B. Definition of the FA-LSP Routing Metric andCreation Cost Function

In current GMPLS standardization [6] there is anintrinsic association between the signaling of new cli-ent LSPs and the creation of the required �-LSP tosupport them. As will be later detailed, a route fromsource to destination is computed upon client LSP re-quest, which may be constituted of both unallocateddata links and already-existent FA-LSPs. In the casethat no FA-LSP is comprised along the route, a new�-LSP is typically set up from source to destination tosupport the incoming request. Otherwise, �-LSPs areset up to provide connectivity on those route segments

Fig. 2. Two-layer network architecture (a) before and (b) after

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

here no FA-LSP is yet established. This operation,owever, may lead to resource waste in the network.otice that long FA-LSPs connecting far-off nodes are

imited to be only reused by incoming LSP requestsetween remote end points. Hence, it appears moreppropriate to separate the signaling functionalityrom the �-LSP creation, so that �-LSP placement cane decided based on network characteristics.

In the present paper, we set the routing metric ofhe already-established FA-LSPs to be max(1, FA-LSPops-1), as described in [6]. Besides, the routing met-ic assigned to the unallocated data links spanningne single physical hop is set to 1. Aiming at better re-ource utilization, however, we dissociate �-LSP es-ablishment from network signaling functionality inhe following way. Once the route from source to des-ination is calculated, the heuristic cost function,

CFA�H� = H��1 − pH� + h/H�, �1�

s applied to the route segments where connectivity isot yet existent, with H standing for the number ofops of the yet-to-be-created �-LSP and pH standing

or the probability that any incoming demand in theetwork has a certain number of hops H. The cost

unction provides us with the most appropriate �-LSPonfiguration to optically connect the yet-uncoveredoute segment. As will be later depicted by example,he term �1−pH� encourages those �-LSP lengths closeo the average network distance, which are thus moreikely to be reused. The term 1/H identifies the use of/E port pairs per hop, so that the larger the �-LSP,

he lower the use of O/E ports to connect its endoints. The tunable h parameter fosters/penalizes the

ting up an end-to-end client LSP supported by three FA-LSPs.

set

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

qduhtmwftB

rpFitFrfi

iinwpoomarps(pFflsa

A262 J. OPT. COMMUN. NETW./VOL. 1, NO. 2 /JULY 2009 Agraz et al.

Au

use of O/E ports in the network. Finally, the total costis multiplied by H as long as �-LSPs need a highernumber of unallocated data links. In this context, letus imagine that a new �-LSP, which will afterwardsact as FA-LSP, has to be established between a nodepair distancing four hops. Supposing that a two-hopclient LSP length is the most likely in the network,the combination CFA�2�+CFA�2�, that is, two �-LSPseach spanning two hops, could have a lower cost thanCFA�4�, meaning one single end-to-end �-LSP. Subsec-tion IV.B particularizes CFA�H� for the scenario understudy. As will be highlighted, very short FA-LSP es-tablishment (e.g., one hop) is also penalized byCFA�H�, due to the large amount of required expensiveO/E ports, as well as the large amount of bypass traf-fic to be electrically processed.

III. CENTRALIZED RESOURCE REALLOCATION MODULE

A. Motivation

As mentioned before, dynamic connection establish-ment along with a randomness of connection holdingtimes may lead to suboptimal resource allocation inthe network at a certain time. For better understand-ing, consider the example depicted in Fig. 3 (left). Itmight happen that, due to the previous resource statein the network, a �-LSP going through A–E–F–G–D–Cwould have to be created to support a client LSP fromnode A to node C. Supposing that the flow requested1/4 of the total wavelength capacity, an FA-LSP fromnodes A to C with 3/4 unreserved bandwidth would becreated (step 1). Imagine now that a client LSP re-quest arrives from node A to node B, also requesting1/4 wavelength capacity. Provided that resourceswould be found on the direct link connecting bothnodes, a direct �-LSP would be set up, resulting aswell in an FA-LSP from node A to C with 3/4 unre-served bandwidth (step 2). Finally, a client LSP re-

Fig. 3. Example of resource reallocation for optim

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

uesting 1/2 of the total capacity reaches node B withestination node C. As the total routing metric of re-sing FA-LSPs from B–A and A–C appears to be muchigher than allocating a direct data link from B to C, ahird �-LSP is set up. This also triggers the establish-ent of a third FA-LSP with 1/2 unreserved band-idth (step 3). Note that, if the client LSP request

rom A to C would have arrived now, rather than someime before, this one would have reused the A–B and–C FA-LSPs.

This arouses concerns for deploying a centralizedesource reallocation module in the network, whicheriodically checks the status of the already-deployedA-LSPs and optimizes client LSP placement accord-

ngly. If a resource reallocation would be triggered inhe situation of Fig. 3 (left), this one could reallocateA-LSP A–C into FA-LSPs A–B and B–C, which wouldesult in resource savings of two client O/E ports andve optical data links as shown in Fig. 3 (right).

In this section, we address the reallocation processn a two-layered transport network. The main targets to minimize both the number of optical resourceseeded to carry the offered traffic to the network, asell as the number of hops of client LSPs. With suchurposes in mind, we first propose an ILP formulationf the problem. Here, our objective lies in finding theptimal solution. In fact, solving the proposed ILP for-ulation is time consuming. Hence, we also introducemetaheuristic to obtain a nearly-optimal resource

eallocation in the network. The optimization processresented in the paper has been called optical re-ources optimization (ORO). Only the client LSPshenceforth paths) are considered in the optimizationrocedures. As a matter of fact, the cost of the existentA-LSPs (hereafter referred to as optical arcs) also re-ects the underlying optical network topology, under-tanding a path as a feasible concatenation of opticalrcs.

tion purposes in a two-layered network scenario.

iza

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

�tsrttwCtsa(rrttcAp

tCciwhihpt

AIOiw

Agraz et al. VOL. 1, NO. 2 /JULY 2009/J. OPT. COMMUN. NETW. A263

Au

B. Optimal Solution: ILP Formulation

In the ILP formulation, the following notations areused for sets and parameters:

E Set of paths (indexed by i)R�i� Set of possible routes for path i (indexed by

j)S Set of optical arcs (indexed by k)Ck Cost of optical arc kMk Capacity of optical arc kNij Equal to 1 if path i was using route j before

optimizationLij Cost of route j for path iQij

k Equal to 1 if route j of path i uses opticalarc k

Wi Bandwidth of path i

Additionally, the following notations are used forvariables:

�k Equal to 1 if optical arc k is used after op-timization (not removed)

�k Optical arc k used bandwidth�ij Equal to 1 if path i uses route j after

optimization�i Equal to 1 if path i has been moved after

optimization�ij Equal to 1 if path i has been moved to route

j after optimization

The ORO procedure releases as much optical re-sources as it can by releasing optical arcs so that theaffected paths are rerouted using the minimum costroute, obtaining the most compact network. The for-mulation is based on an arc-path model, where the setof distinct routes for every path must be precomputed.Hence, the following objective function:Minimize

− �1 � �k�S

�Ck � �1 − �k��

+ �2 � �i�E

�j�R�i�

��ij � Lij� + �3 � �k�S

�k �2�

subject to:

�j�R�i�

�ij = 1, ∀ i � E, �3�

�i�E

�j�R�i�

Wi � Qijk � �ij �k, ∀ k � S, �4�

�k �k �Mk � �k�, ∀ k � S, �5�

�i +�k�S��k � �j�R�i��Nij � Qij

k��

�k�S�j�R�i��Nij � Qijk�

1, ∀ i � E,

�6�

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

�k�S

��1 − �k� � �j�R�i�

�Nij � Qijk�� − �i 0, ∀ i � E,

�7�

�ij = �ij − Nij � �ij, ∀ i � E,j � R�i�, �8�

�j�R�i�

�ij = �i, ∀ i � E, �9�

�k,�ij,�i,�ij � �0,1, �k integer. �10�

In the formulation, parameters �1, �2, and �3 ��1�2��3� in multiobjective function (2) give focus on

he optical arc reduction objective. Constraint (3) en-ures that every path has one and only one assignedoute. Constraints (4) and (5) guarantee that if an op-ical arc is removed from the network, it does not con-ain any path. Moreover, the optical arc used band-idth is stored and capacity restrictions are checked.onstraint (6) makes sure that each path using an op-

ical arc to be removed is rerouted. Conversely, con-traint (7) ensures that paths supported on opticalrcs kept in the solution are not rerouted. Constraint8) stores the new route to be used by the path beingerouted. Constraint (9) provides the paths to be re-outed. Note that constraints (8) and (9) basically cap-ure information about the path routes, simplifyinghe structure of the objective function as well. Finally,onstraint (10) defines variables as binary or integer.dditionally, we assume that every optical arc sup-orts at least one path.

The formulation was run using the ILOG OPL [17]o implement the model and solved using the ILOGPLEX v.11.0 [17] optimizer on a 3 GHz CPU ma-

hine with 1 GB RAM memory. Being that the models quite complex, we observed in some example net-ork scenarios that the solution lasted more than oneour. To enable ORO application in an off-line central-

zed module in the network, we also provide a meta-euristic based on greedy randomized adaptive searchrocedures (GRASP) [18] aiming at lower response op-imization time.

lgorithm 1: ORO Constructive Phasenput: Candidate list, network graphutput: List of optical arcs to remove, list of paths to move with

ts new route and the cost of the solutionhile candidate list size�0 doRCL= �optical arc�candidate list:BwBwmin

+��Bwmax−Bwmin�;Get a random optical arc from the RCL and remove it from thecandidate list;if reroute (arc, network graph) then

Add the optical arc to the solution;Add every moved path to the solution;If there is any path in its original route, remove it from thesolution;

Compute solution cost;

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

nonbti

apAfapcit

sciattpOpitp

A

ocnset1docltinpn

tOR

A264 J. OPT. COMMUN. NETW./VOL. 1, NO. 2 /JULY 2009 Agraz et al.

Au

Algorithm 2: ORO Reroute AlgorithmInput: Arc to reroute, network graphOutput: Successful or failed rerouteCreate an empty path list;Remove the arc from the graph;for every path using the optical arc do

Get a copy of the current route of the path;Tear down the path;Compute shortest path from source to destination;if no route found then

Restore the original route of the path and set it up;Add the optical arc to the graph;for every path in the list do

Tear down the path;Establish the path in its original route;

Return (failed to reroute);Establish the path;Add the path to the list;

Mark every path in the list as moved;Return (success);

C. Proposed Heuristics

The ORO metaheuristic constructive phase (Algo-rithm 1) consists of trying to remove optical arcs fromthe network graph. At the beginning, a candidate listis built, containing every optical arc in the graphsorted in increasing order by used bandwidth �Bw�.For each iteration, an optical arc is selected from a re-stricted candidate list (RCL), which only contains asubset of the total candidate list composed of theleast-congested optical arcs. In operation, the algo-rithm randomly chooses an optical arc and subse-quently removes it from the candidate list. Then, ittries to reroute all paths supported on this selectedoptical arc.

The ORO reroute algorithm (Algorithm 2) uses theDijkstra shortest path algorithm [19] to find an alter-native route from source to destination, particularlyavoiding the optical arc to be removed and withenough bandwidth to support the rerouted path. Notethat the Dijkstra algorithm can only be used on con-nected graphs with no parallel arcs. This is not neces-sarily accomplished by the client aggregation layer asa whole. Because connections are dynamic in nature,the graph may not be connected, appearing as severalconnected subgraphs. This is illustrated in the ex-ample shown in Fig. 3. There, it can be realized thatparallel arcs are easily found. In fact, the nodal degreed in the client layer is typically much higher than inthe optical layer, as can be appreciated in node Dwhere d equals 5. For instance, if a DWDM opticallayer with 40 wavelengths per link and an averagenode degree d̄=3 would have been considered, a d̄value equal to 120 could be reached in the client layer.Hence, before running the construction algorithm, avirtual node is added when a parallel arc between two

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

odes is found. Besides, the Dijkstra algorithm is runver the connected subgraph containing the originode. In the case that a path in the optical arc cannote rerouted, already-processed paths are returned tohe original route and the optical arc is placed againn the network graph.

The constructive phase analyzes all optical arcs �m�nd for each one it tries to reroute every supportedath on the arc. Let us consider a worst-case scenario.s mentioned above, if no alternative route would be

ound for the last path under the reroute process, alllready-rerouted paths contained in the arc would belaced again in their original route. This results in aomputational complexity of O�m�p2�, with p stand-ng for the maximum number of paths contained inhe optical arcs.

The ORO metaheuristic improvement phase con-ists of optical arc exchanges. An optical arc in theurrent solution is returned to the network and themprovement algorithm tries to get alternative opticalrcs where paths may be able to be rerouted. Beinghat the exchange cost is lower than the cost of the re-urned arc, the current solution is updated. The com-utational complexity can be demonstrated to be�m2�p2�. Because the set of construction and im-rovement phases is repeated k times, updating thencumbent solution when the cost of the current solu-ion is lower than the cost stored so far, the total com-utational complexity rises to O�k�m2�p2�.

IV. EXPERIMENTAL RESULTS

. ASON/GMPLS CARISMA Test Bed

The experimental evaluation has been carried outver the ASON/GMPLS CARISMA test bed [20], aonfigurable multitopology signalling communicationsetwork (SCN) running over wavelength-selective-witch- (WSS-) based optical cross-connect (OXC)mulators. In this configurable SCN, optical connec-ion controllers (OCCs) are interconnected by00 Mbps full-duplex point-to-point Ethernet links,escribing the same physical topology of the emulatedptical transport plane. This results in an out-of-fiberontrol plane architecture associated with the under-ying transport plane. Running on top of the architec-ure, a network management system (NMS) has beenmplemented as a web-based application, enablingetwork configuration and both permanent and soft-ermanent connection provisioning through the Inter-et.

In the test bed, OCCs are deployed by means of Pen-ium IV Linux-based routers at 2 GHz, so that eachCC implements the full GMPLS protocol set:SVP-TE for signaling, OSPF-TE for routing and in-

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

aaLeut

lstmadllqacRcmuaPp�

cRbLeeoptP

ctcIcFAa

CE

oppga

Agraz et al. VOL. 1, NO. 2 /JULY 2009/J. OPT. COMMUN. NETW. A265

Au

formation advertisement, and, finally, link manage-ment protocol (LMP) for resource discovery and man-agement [21]. In this work, we particularly build anine-node meshed SCN configuration to evaluate theproposed contributions as depicted in Fig. 4.

B. Forwarding Adjacencies Implementation

Because the CARISMA test bed was essentially awavelength-routed optical network, resources were al-located for end-to-end connections with a whole wave-length granularity. This section reports the implemen-tation and further evaluation of the entire FA-LSPfunctionality in the CARISMA test bed, enhancing itwith multilayer support at the control plane level.

Looking at the standardization [22], a newRSVP-TE object named theLSP�TUNNEL�INTERFACE�ID was proposed to beused when signaling a new FA-LSP in the PATH andRESV RSVP-TE messages [9]. The object contains twofields, that is, the FA-LSP identifier and the router ID.

In the event of a new �-LSP to be set up, the head-end OCC must allocate an identifier for the interfaceassociated with the yet-to-be-created FA-LSP. Next, itoriginates an RSVP-TE PATH message containing aLSP�TUNNEL�INTERFACE�ID object filled with theselected local interface identifier, along with the localoptical node identifier. When the PATH message ar-rives at the destination, the tail-end OCC must allo-cate an identifier for that FA-LSP end. This is calledthe remote FA-LSP interface identifier, which is re-ported back to the head end within the RSVP-TERESV message. As soon as the �-LSP has been cre-

Fig. 4. The ASON/GMPLS CARISMA test bed describing a nine-node meshed scenario.

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

ted, the head-end OCC advertises it as a forwardingdjacency by means of OSPF-TE. Being that the FA-SP is bidirectional, it is also advertised by the tail-nd OCC. All OCCs receiving the FA advertisementpdate their link state database adding a new link be-ween the involved nodes.

Note that in the GMPLS standardization, the estab-ishment of an FA-LSP is intrinsically associated withetting up a client layer LSP. As introduced in Subsec-ion II.B, however, �-LSP length may be limited toaximize resource reutilization in the network. To

chieve such purposes, �-LSP establishment must beissociated from the client LSP setup procedure, al-owing in this way the establishment of several under-ying �-LSPs while signaling only one client LSP re-uest. To permit this separation between client LSPnd �-LSP setup, the head-end OCC evaluates the ac-umulated optical length, using loose hops in theSVP-TE explicit route object (ERO) [22] when it de-

ides to divide the whole route segment into two orore underlying �-LSPs. The same mechanism is

sed also when an intermediate �-LSP must be cre-ted. Every intermediate OCC receiving an RSVP-TEATH message with a next hop set as loose must com-ute the next route segment possibly signaling a new-LSP.

It is worth pointing out that �-LSPs created duringlient LSP setup signaling procedures have no specificSVP-TE refresh messages. In fact, the �-LSP muste torn down when releasing the last supported clientSP. In order to maintain the association between cli-nt LSPs and �-LSP everywhere, the signaling of cli-nt LSPs using an already-existing FA-LSP follows, inur implementation, the same path in the controllane as the signal at the optical layer. This meanshat the same OCC can process a single RSVP-TEATH or RESV message more than once.

In the example in Fig. 5, nodes E and D are directlyonnected through FA-LSP E–D. Let us suppose thathe route of a new client LSP between nodes A–G in-ludes FA-LSP E–D together with A–E and D–G links.n such a case, two new FA-LSPs A–E and D–G arereated to support the new client LSP A–G. Node G inig. 5 processes RSVP-TE messages of the client LSP–G two times, that is, the first time as an intermedi-te OCC and the second time as the tail-end OCC.

. Scenario Configuration and FA-LSP Functionalityvaluation

For the evaluation, we assume a nine-node meshedptical network with the same topology as the controllane depicted in Fig. 4 (i.e., the configured controllane would be deployed associated with a nine-noderooming-capable optical transport network). In suchscenario, we assume that each link carries eight bi-

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

t5Twib(nfim+

pw

Focca

A266 J. OPT. COMMUN. NETW./VOL. 1, NO. 2 /JULY 2009 Agraz et al.

Au

directional wavelengths. For the traffic characteris-tics, we consider that uniformly distributed client LSPrequests arrive at the network following a Poissonprocess with mean interarrival time (IAT) equal to1/�. Besides, connection duration follows an exponen-tial distribution with mean holding time (HT) set to1/�. In particular, the requested Bw of all incomingclient LSP requests is considered to be 1/4 of the totalwavelength capacity.

Figure 6 illustrates the obtained CFA�H� function forthe scenario under study. The bar graph plots theprobability that an incoming client LSP request has acertain number of hops, assuming availability of re-sources through the shortest path. As seen, there is a40% probability that an incoming request traverses

Fig. 5. Example of

Fig. 6. (Color online) FA-LSP creation cost function for the nine-node topology under study. Bar graph shows the probability that aclient LSP request has 1, 2, 3, or 4 hops.

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

wo hops from source to destination. In contrast, only% of the incoming requests would traverse four hops.his validates our assumption in Subsection II.B,here we stated that, by splitting very long FA-LSPs

nto shorter ones, resources are much more likely toe reused. Values greater than the network diameteri.e., four hops in our scenario) have pH=0.0 and areot depicted in the figure. To finally obtain CFA�H� wex h=0.5, as it provided the best network perfor-ance, while fulfilling our design criteria: CFA�2�CFA�2� CFA�4� and CFA�1�+CFA�2� CFA�3�.

Figure 7 plots the obtained client LSP blockingrobability as a function of the offered load to the net-ork �� /�=HT/IAT� normalized to a value of 200.

LSP establishment.

ig. 7. (Color online) Client LSP blocking probability as a functionf the offered load to the network. The situations without FA-LSPapability in the network, with FA-LSP capability, and with FA-LSPapability also deploying a centralized resource reallocation modulere considered.

FA-

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

priittsbnrrwbr0rTw

ftvfibnabnedrr

ap

Fa

Agraz et al. VOL. 1, NO. 2 /JULY 2009/J. OPT. COMMUN. NETW. A267

Au

The curve without FA-LSP indicates the situationwhere no FA-LSP capabilities exist in the network,thus allocating a whole wavelength capacity for theincoming connection requests (even though they onlyrequest 1/4 of the total wavelength capacity). Con-versely, the curve with FA-LSP evaluates the improve-ment obtained when enhanced subwavelength provi-sioning flexibility is provided to the network byimplementing GMPLS-controlled grooming actions(i.e., with FA-LSP). The situation of additionally de-ploying a centralized resource reallocation module inthe network is also depicted, but it will be detailed inthe following subsection.

As expected, significantly better resource usage isachieved when implementing FA-LSP capabilities inthe network. For instance, if 1% client LSP blockingprobability would have to be ensured, a maximumload of L=0.05 could be offered to a pure wavelength-routed optical network (i.e, without FA-LSP). Con-versely, it could be further increased to L=0.55 whenFA-LSPs are implemented. This �L1=0.5 experimen-tally assesses the implementation of FA-LSP capabili-ties to automatically manage grooming actions in fu-ture transport networks, given the lack of purewavelength-routed optical networks to allocate incom-ing sub-lambda client LSP requests. In fact, as awhole wavelength is allocated in a per 1/4 wavelengthcapacity client LSP request in the without FA-LSPsituation, 3/4 of the total network capacity is directlywasted. Qualitatively speaking, this approximatelyresults in four times less carried traffic by the net-work.

D. Deployment and Validation of the CentralizedResource Reallocation Module

On the basis of an already-operative GMPLS-controlled multilayer network, a centralized resourcereallocation module has been further deployed andevaluated in the CARISMA test bed. First of all, theperformance of the ORO metaheuristic has been com-pared to the optimal solution obtained with the OROILP formulation. In the situations depicted in Fig. 7low relative errors to the optimal solution around 5%are observed for low- and medium-loaded network sce-narios (i.e., leading to client LSP request blockingprobabilities lower than 1%). In fact, relative errorswere significantly increased to 10%–20% in highlyloaded network situations. Note, however, that theseloads lead to unacceptable blocking probability valueshigher than 5%. Furthermore, it is worth highlightingthat measures showed less than 0.3 s ORO metaheu-ristic running times. These short running times lever-age its applicability for optimizing resource allocationin the network.

Because the client LSP requests incoming process

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

lays an important role in final network performance,esource reallocation actions are triggered every 50ncoming requests. In our scenario this is translatednto applying the resource reallocation process a fewimes per day. Looking at the results previously ob-ained in Fig. 7, a reduction in the number of con-umed resources and its consequent impact on thelocking probability can be appreciated. The randomature of the offered connections leads to suboptimalesource utilization, so the proposed periodic rear-angement module is useful to improve the whole net-ork performance. When looking at the figure, it cane noticed that request blocking probability is clearlyeduced for offered load values ranging from 0.4 to.65. For example, the request blocking probability iseduced from 2% to around 1% when the load is 0.625.his is due to the release of some O/E ports producedhen reallocating client LSPs between FA-LSPs.

Port reduction is additionally plotted in Fig. 8 as aunction of offered network load. There, it is shownhat for low load values the gain is only marginal. Ob-iously, when the number of FA-LSPs is low, it is dif-cult to reallocate traffic. As a matter of fact, OROenefits increase along with the offered load to theetwork, reaching its maximum for load valuesround 0.6. Under these loads, 14% port reduction cane achieved in the network, drastically decreasingetwork capital expenditures (CAPEX). Note, how-ver, that if the load is further increased, reductionecreases again as the FA-LSP mean occupationeaches higher values, it thus being again difficult toeallocate client LSPs.

V. CONCLUDING REMARKS

This paper targeted at designing and implementingGMPLS-controlled grooming-capable optical trans-

ort network. To this end, the current standardization

0 0.125 0.25 0.375 0.5 0.625 0.75 0.875 1Offered load

0

2

4

6

8

10

12

14

16

Por

tred

uctio

n(%

)

Port reduction

ig. 8. (Color online) O/E port reduction in the network in percent-ges when the ORO metaheuristic is applied in the network.

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

n

A268 J. OPT. COMMUN. NETW./VOL. 1, NO. 2 /JULY 2009 Agraz et al.

Au

framework concerning GMPLS-enabled multilayernetworks was first reviewed and exemplified. Next,operation of a centralized resource reallocation mod-ule to be deployed in a multilayer optical network wasdiscussed. In this context, an ILP formulation quanti-fying its optimal performance was derived. Being thatthe model is quite complex and its solving time israther long, subsecond running time metaheuristicswere also provided, which obtained less than 5% rela-tive error to the optimal solution in the operating net-work scenario. The obtained experimental resultsvalidated GMPLS-controlled grooming actions, drasti-cally improving client LSP blocking probability com-pared to a pure wavelength-routed optical networkscenario. Besides, the deployment of the centralizedresource reallocation module in the ASON/GMPLSCARISMA test bed resulted in noticeable improve-ments in terms of both O/E port and client LSP block-ing probability reduction. In particular, focusing on ascenario with request blocking probabilities around1%, O/E port and request blocking probability reduc-tions of about 10% and up to 1% were respectively ob-tained.

The evaluation presented in this paper concerns asingle nine-node ASON domain. Further work will ex-tend the implementation of the FA-LSP functionalityin larger network scenarios, including signaling androuting interworking issues in multidomainmultilayer network environments. Particularly, it isour goal to assess not only the performance, but alsothe scalability of the GMPLS-controlled grooming andthe proposed reallocation approach as the networkgets larger, even spanning more than a single domain.

ACKNOWLEDGMENT

The work reported in this paper has been partiallysupported by the Spanish Science Ministry throughthe project Engineering Next Generation OpticalTransport NEtworks (ENGINE), (TEC2008-02634).Moreover, the authors thank the support from thei2CAT Foundation through the project TRILOGY.

REFERENCES

[1] ITU-T Rec. G.8080/Y.1304, “Architecture for the automaticallyswitched optical networks,” Nov. 2001.

[2] E. Mannie, “Generalized multi-protocol label switching (GM-PLS) architecture,” IETF RFC 3945, Oct. 2004.

[3] O. Gerstel, P. Lin, and G. Sasaki, “Combined WDM and SO-NET network design,” in Proc. IEEE INFOCOM, 1998, pp.734–743.

[4] K. Zhu and B. Mukherjee, “Traffic grooming in an optical WDMmeshed network,” IEEE J. Sel. Areas Commun., vol. 20, no. 1,pp. 122–133, Jan. 2002.

[5] R. Dutta and G. N. Rouskas, “Traffic grooming in WDM net-works: past and future,” IEEE Network, vol. 16, no. 6, pp. 46–56, Nov. 2002.

[6] K. Kompella and Y. Rekhter, “Label switched paths (LSP) hi-

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

erarchy with generalized multi-protocol label switching (GM-PLS) traffic engineering (TE),” IETF RFC 4206, Oct. 2005.

[7] J. Comellas, R. Martínez, J. Prat, V. Sales, and G. Junyent,“Integrated IP/WDM routing in GMPLS-based optical net-works,” IEEE Network, vol. 17, no. 2, pp. 22–27, Mar. 2003.

[8] D. Katz, K. Kompella, and D. Yeung, “Traffic engineering (TE)extensions to OSPF version 2,” IETF RFC 3630, Sept. 2003.

[9] L. Berger, “Generalized multi-protocol label switching (GM-PLS) signaling resource reservation protocol-traffic engineer-ing (RSVP-TE) extensions,” IETF RFC 3473, Jan. 2003.

[10] D. Awduche, L. Berger, D. Gan, T. Li, V. Srinivasan, and G.Swallow, “RSVP-TE: extensions to RSVP for LSP tunnels,”IETF RFC 3209, Dec. 2001.

[11] B. Ramamurthy and A. Ramakrishnan, “Virtual topology re-configuration of wavelength routed optical WDM networks,” inProc. IEEE GLOBECOM, 2000, pp. 1269–1275.

[12] R. Mahalati and R. Dutta, “Reconfiguration of traffic groomingoptical networks,” in Proc. BROADNETS, 2004, pp. 170–179.

[13] A. Gencata and B. Mukherjee, “Virtual-topology adaptation forWDM mesh networks under dynamic traffic,” IEEE/ACMTrans. Netw., vol. 11, no. 2, pp. 236–247, April 2003.

[14] ITU-T Rec. G.805, “Generic functional architecture of trans-port networks,” March 2000.

[15] D. Fedyk, O. Aboul-Magd, D. Brungard, J. Lang, and D. Pa-padimitriou, “A transport network view of the link manage-ment protocol (LMP),” IETF RFC 4394, Feb. 2006.

[16] I. Bryskin and A. Farrell, “A lexicography for the interpreta-tion of generalized multi-protocol label switching (GMPLS) ter-minology within the context of the ITU-T’s automaticallyswitched optical network (ASON) architecture,” IETF RFC4397.

[17] ILOG Inc., www.ilog.com[18] T. Feo and M. Resende, “Greedy randomized adaptive search

procedures,” J. Glob. Optim., vol. 6, pp. 109–133, June 1995.[19] R. Bhandari, Survivable Networks: Algorithms for Diverse

Routing, Norwell, MA: Kluwer Academic, 1999.[20] J. Perelló, E. Escalona, S. Spadaro, J. Comellas, and G. Juny-

ent, “Resource discovery in ASON/GMPLS transport net-works,” IEEE Commun. Mag., vol. 45, no. 8, pp. 86–92, Aug.2007.

[21] J. Lang, “Link management protocol (LMP),” IETF RFC 4204,Oct. 2005.

[22] K. Kompella and Y. Rekhter, “Signalling unnumbered links inresource reservation protocol—traffic engineering (RSVP-TE),”IETF RFC 3477, Jan. 2003.

Fernando Agraz received his M.Sc. degreein computer engineering in 2005 from thePolytechnic University of Catalonia (UPC).Since 2005 he has been working as a re-search engineer in the Optical Communica-tions Group (GCO) at UPC, also preparinghis Ph.D. He has also participated in vari-ous European research projects such asIST Nobel Phase 2 or E-Photon/ONe�.His current research focuses on networkmanagement and routing in GMPLS-based

etworks.

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.

CtaHe

aa

CA(

Agraz et al. VOL. 1, NO. 2 /JULY 2009/J. OPT. COMMUN. NETW. A269

Au

Luis Velasco received the B.Sc. degree intelecommunications engineering from Uni-versidad Politécnica de Madrid (UPM) in1989, the M.Sc. degree in physics from Uni-versidad Complutense de Madrid (UCM) in1993, and the Ph.D. degree from Univer-sitat Politècnica de Catalunya (UPC) in2009. In 1989 he joined Telefónica of Spainand was involved in the specifications andfirst office application of the Telefónica SDHtransport network. In 2003 he joined UPC,

where currently he is an Assistant Professor in the Computer Ar-chitecture Department (DAC) and a Researcher in the Optical Com-munications group (GCO) and the Advanced Broadband Communi-cations Center (CCABA). His interests include signaling, routing,and resilience mechanisms in ASON/GMPLS-based networks.

Jordi Perelló received his M.Sc. degree intelecommunications engineering in 2005from the Universitat Politècnica de Catalu-nya (UPC). Currently, he is an AssistantProfessor in the Computer Architecture De-partment (DAC) at UPC, finishing his Ph.D.studies in the Advanced Broadband Com-munications Center (CCABA). He has par-ticipated in various IST FP-6 and FP-7 Eu-ropean research projects such as EUDICONET, BONE, IST NOBEL 2, e-Photon/

ONe�, and COST Action 291. His research interests concern re-source management, quality of service issues, and survivability ofnext-generation optical transport networks.

Marc Ruiz received the B.Sc. degree in bi-ology from Universitat de Barcelona (UB),Spain, in 2005 and the M.Sc. degree in sta-tistics and operational research from Uni-versitat Politècnica de Catalunya (UPC),Barcelona, Spain, in 2009. He is currentlyworking towards the Ph.D. degree with theOptical Communication Group (UPC). Hisresearch interests include optimization andperformance of next-generation optical net-works.

thorized licensed use limited to: UNIVERSITAT POLIT?CNICA DE CATALUNYA. Dow

Salvatore Spadaro received the M.Sc.(2000) and the Ph.D. (2005) degrees in tele-communications engineering from Univer-sitat Politècnica de Catalunya (UPC). Healso received the Dr.Ing. degree in electricalengineering from Politecnico di Torino(2000). He is currently an Associate Profes-sor in the Optical Communications Group ofthe Signal Theory and Communications De-partment of UPC. Since 2000 he has been astaff member of the Advanced Broadband

ommunications Center (CCABA) of UPC, and he is currently par-icipating in the DICONET and BONE FP7 EU projects. He has co-uthored about 80 papers in international journals and conferences.is research interests are in the field of all-optical networks with

mphasis on traffic engineering and resilience.

Gabriel Junyent is a telecommunicationsengineer (Universidad Politécnica deMadrid, UPM, 1973) and holds a Ph.D. de-gree in communications (UPC, 1979). Hehas been a Teaching Assistant (UPC, 1973–1977), Adjunct Professor (UPC, 1977–1983),Associate Professor (UPC, 1983–1985), andProfessor (UPC, 1985–1989) and has been aFull Professor since 1989. In the past15 years he has participated in more than30 national and international R&D projects

nd has published more than 30 journal papers and book chaptersnd 100 conference papers.

Jaume Comellas received the M.S. (1993)and Ph.D. (1999) degrees in telecommunica-tions engineering from UPC. His current re-search interests are optical transmissionand IP over WDM networking topics. Hehas participated in many research projectsfunded by the Spanish government and theEuropean Commission. He has co-authoredmore than 70 research articles in interna-tional journals and conferences. He is an as-sociate professor in the Signal Theory and

ommunications Department at UPC, also serving as Internationalffairs Vice-Dean at the Telecommunications Engineering School

Telecom BCN) of the same university.

nloaded on June 08,2010 at 10:47:00 UTC from IEEE Xplore. Restrictions apply.