ew

iore

ordi Ga 29,126 N

Article history:Received 30 September 2011

the ecological footprint of the network elements (NEs) isexplicitly taken into account. In this scenario, governmentsand society are endorsing the development of greenrenewable energy sources (such as solar panels, wind

Recent studies [1] conrm that the use of opticaltechnology in high-capacity switches and routers is moreenergy-efcient than electronic technology and that cir-cuit-switched architectures consume signicantly lessthan their packet-switched counterparts. However, despitethe recent efforts in improving the energy-efciency of theinvolved technological components [2], the amount ofpower to be spent worldwide for powering network

1389-1286/$ - see front matter 2012 Elsevier B.V. All rights reserved.

Corresponding author.E-mail addresses: sergior@ac.upc.edu (S. Ricciardi), fpalmier@unina.it

(F. Palmieri), uore@unina.it (U. Fiore).

Computer Networks 56 (2012) 24202442

Contents lists available at SciVerse ScienceDirect

Computer N

.e lshttp://dx.doi.org/10.1016/j.comnet.2012.03.0161. Introduction

The containment of power consumption and the reduc-tion of the associated green house gases (GHG, mainly CO2)emissions are emerging as new challenges for telecommu-nication operators. In fact, the rising energy costs due tothe scarcity of fossil fuels, the increasingly rigid environ-mental standards and the growing power requirementsof modern high-performance networking devices areimposing new constraints, further stressing the require-ments towards an energy-aware business model in which

turbines, and geothermal plants) for powering NEs. Greenenergy sources are preferable with respect to the tradi-tional dirty ones (e.g. coal, fuel, gas) since they do notemit GHG in the atmosphere while producing electricalenergy. Nonetheless, green energy sources (e.g. wind,sun, tide) are not always available at all sites and arevariable with time; for this reason, the NEs that are pow-ered by green energy sources are also provided with thelegacy, dirty sources. At the occurrence, the smart gridpower distribution system switches to the dirty powersupply without any energy interruption.Received in revised form 23 February 2012Accepted 20 March 2012Available online 29 March 2012

Keywords:Green networksCross-layer optimizationsEnergy consumptionGHG emissionsPower demand in networking equipment is expected to become a main limiting factor andhence a fundamental challenge to ensure bandwidth scaling in the next generation Inter-net. Environmental effects of human activities, such as CO2 emissions and the consequentglobal warming have risen as one of the major issue for the ICT sector and for the society.Therefore, it is not surprising that telecom operators are devoting much of their efforts tothe reduction of energy consumption and of the related CO2 emissions of their networkinfrastructures. In this work, we present a novel integrated routing and wavelength assign-ment framework that, while addressing the traditional network management objectives,introduces energy-awareness in its decision process to contain the power consumptionof the underlying network infrastructure and make use of green energy sources whereverpossible. This approach results in direct power, cost and CO2 emissions savings in the shortterm, as demonstrated by our extensive simulation studies.

2012 Elsevier B.V. All rights reserved.a r t i c l e i n f o a b s t r a c tAn energy-aware dynamic RWA framwavelength-routed networks

Sergio Ricciardi a,, Francesco Palmieri b, Ugo FJosep Sol-Pareta a

aDepartment of Computer Architecture, Technical University of Catalonia, c. JbDepartment of Information Engineering, Second University of Naples, v. RomcCenter of Information Services, University of Naples Federico II, v. Cinthia, 80

journal homepage: wwwork for next-generation

c, Davide Careglio a, Germn Santos-Boada a,

irona 1-3, 08034 Barcelona, Spain81031 Aversa, Italyaples, Italy

etworks

evier .com/locate /comnet

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2421infrastructures can be globally quantied in the order oftens of gigawatts, corresponding to more than 1% of theworldwide electricity consumption [2] (to give an idea,the equivalent of 22 nuclear reactors are needed to gener-ate such a huge power demand). Thus, limiting power con-sumption in network infrastructures can bring greatbenets and reduce their overall ecological footprint, sothat the need for a greener, energy-aware Internet is rap-idly becoming a fundamental political, social and commer-cial issue. Furthermore, with the ever increasing demandfor bandwidth, connection quality and end-to-end interac-tivity, computer networks are requiring more and moresophisticated and power-hungry devices, such as high-end routers, signal regenerators, optical ampliers, recon-gurable add-and-drop multiplexers and very fast (digitalsignal) processing units. These components tend to in-crease the energy needs of global communication expo-nentially so that power consumption is becoming asignicant limiting factor for the overall scalability ofnext-generation high-capacity telecommunication net-works. In the next years, large-scale optical transport infra-structures will no longer be constrained mainly by theircapacity, but rather by their energy consumption costsand environmental effects [3].

As a consequence, it is necessary to envisage how thenext-generation network architectures and protocols canbe modied to meet the purpose of energy-efciency.Unfortunately, the rush for achieving energy-efciencyresulted in the fact that many of the solutions proposedto-date (e.g. [4,5]) tend to minimize only the energy con-sumption of the networks while disregarding the tradi-tional network management goals such as the overallnetwork load-balancing. It is instead mandatory to guaran-tee that the above modications will not adversely affectthe fundamental operators optimization objectives ofkeeping the resource usage fairly balanced, to save on eachavailable link sufcient free capacity for demands that mayreasonably emerge in the infrastructure operating lifetime,and minimizing the network usage costs, considered as astatic way of expressing operator preference to choosesome favorite link resources. In the ideal case, new solu-tions should not only lower the ecological footprint, butalso increase the offered quality-of-service such as theconnection blocking probability.

Starting from the above considerations, we introduce[6] energy-awareness into control plane protocols whosegoal is to properly condition the route/path selectionmechanisms on relatively coarse time scales by privilegingthe use of green energy sources and energy-efcient links/switching devices, simultaneously taking advantage fromthe different users demands across the interested networkinfrastructures. The selected paths are likely not to be theshortest or best ones, but the resulting power and GHGsavings are substantial, and possible losses on the otheroptimization objectives (i.e. number of blocked connectionrequests) are taken into account and kept as low as possi-ble. In such a way, the overall power consumption andGHG emissions can be minimized while the traditionaloptimization objectives (such as load-balancing) are notdisrupted. In doing this, we combine all the notable fea-tures that a comprehensive energy-aware network modelshould have and put them together into a general routingand wavelength assignment (RWA) framework.

The RWA problem is known to be NP-complete [7] andin the dynamic case no optimality is possible since there isno previous knowledge of the connection requests that willbe handled by the network. Therefore, we introduce a newheuristic method for efciently calculating (in polynomialtime) the routing information subject to power consump-tion constraints, taking into account also the specic kindof energy source (dirty or green) used for powering thetraversed NEs. In order to evaluate the performance ofour approach, we compared our approach with well-known RWA algorithms in the literature. The proposedapproach, that in the following will be referred to as Green-Spark, introduces energy-awareness into the Spark frame-work [8], which is a two-stage integrated RWA schemestructured in a pre-selection phase where a number of kcandidate paths satisfying the connection constraints arefound and a nal selection stage where the optimum pathamong the candidates determined in the previous phase ischosen according to a properly crafted heuristic. Green-Spark is a simple and effective two-stage on-line RWAscheme providing wavelength routing as well as groomingcapabilities in the state-of-the-art hybrid electric-opticalnetwork infrastructure. In its rst stage, this enhancedRWA scheme nds, for each new connection request, aset of feasible lightpaths satisfying both the users specicend-to-end demands (QoS, bandwidth, etc.) and traditionaloptimization objectives, while in the second stage it basesits nal choice on the aforementioned power and GHG con-tainment requirements. In the end, it nally achieves anoptimal trade-off between energy optimization and net-work/users requirements in an affordable computationaltime. GreenSpark differs from Spark for the second stage,which has been here introduced to meet the energy-re-lated criteria. Furthermore, Spark used a special parameter(kHop) to explicitly limit the length of lightpaths, whilst inGreenSpark this is not needed anymore due to the additivenature of the energy consumption function: longer pathswill have higher energy consumption and, thus, will havelower probability to be chosen for the routing of theconnections.

GreenSpark is based on a totally exible network modelsupporting heterogeneous equipment, in which the num-ber and type of lambdas can vary on each link or node,together with the associated power consumption, and pro-vides a fully dynamic path selection scheme in which thegrooming policy is not predetermined but may vary, alongwith the evolution of the network trafc. We explicitlyconsider the inuence of trafc on power consumptionby using realistic data for trafc demands, network topol-ogies, link costs, and energy requirements of the NEs.

This approach is also based on deeper network engi-neering considerations that make it behave very differentlyfrom the other already existing energy-aware networkingapproaches, mainly based on the concept of temporarilyswitching off entire devices or subsystems (the least usedones) in order to minimize energy consumption by rerout-ing the involved trafc. Such approaches, often referred assleep mode [9], may be unpractical, especially for large andhighly connected switching nodes, since many very

sented by Idzikowski et al. [5]. They analyzed the effects of

Most of these approaches are characterized by a limited

grooming capability is required on opto-electronic routers

2422 S. Ricciardi et al. / Computer Networks 56 (2012) 24202442reconguring routing, at the different layers, by assumingcomplete wavelength conversion capability in each node.In [20], Chiaraviglio et al. have proposed and evaluatedsome greedy heuristics based on the ranking of nodesand links with respect to the amount of trafc that theywould carry in the context of an energy-agnostic congu-ration. Silvestri et al. [21] combined trafc grooming andtransmission optimization techniques to limit energy con-sumption in the WDM layer. Trafc grooming shifts trafcfrom some links to other ones in order to switch emptyones off, and transmission optimization adjusts dispersionmanagement and pulse duration, which decreases theneed for using in-line 3R regenerators. The power savingsthat can be achieved by dynamically adapting the networktopology to the trafc volume are investigated in [22],expensive transmission links become unused, hence leav-ing signicant capital investments (CAPEX) unproductivefor the entire duration of the sleep interval. Furthermore,sleep mode drastically reduces the overall meshing degree,by limiting the network reliability, and partially negatesthe possibility of balancing the load on multiple availablelinks/paths [10]. Finally, results in [11] show that sleepmode is achievable just for very few nodes and only at verylow loads. Conversely, in our model, energy-aware archi-tectures allow the NEs power consumption to scale withtrafc load, as in [1014]; such architectures are stronglyadvocated by current efforts from standardization bodiesand governmental programs [15] and can be made upusing off-the-shelf standard technologies [16,17].

2. Related work

Greening the network is an active subject of recentresearch. Several papers have concentrated on the reduc-tion of power consumption. In [13] Gupta and Singh wereamong the rst researchers to envision the idea of energyconservation in Internet-based infrastructures. Shen andTucker [4] developed mixed integer linear programming(MILP) methods and heuristics to optimize the energy con-sumption of a IP over WDM transport network. In detail,their objective was minimizing power consumption ofthe network by switching off router ports, transponders,and optical ampliers; they proposed two heuristics (di-rect bypass and multi-hop bypass). Another approachfocusing on a MILP-based formalization of the power-aware routing and wavelength assignment has been alsopresented by Wu et al. [18]. In their work, energy savingscan be achieved by switching off optical cross connects(OXC) and optical ampliers according to three differentalgorithms and criteria. In [19,10], ILP mathematical for-mulations are presented with the double objective ofreducing both the energy consumption and the GHG emis-sions of network infrastructure. Since the ILP solves at theoptimum the ofine RWA problem, these works give anupper bound for energy and GHG savings. However, usingILP for real-world networks with dynamic trafc is unprac-tical due to its intractable computational complexity. En-ergy saving by dynamically switching off idle IP routerline cards in low-demand hours was also the approach pre-operating on the network edge. Accordingly, all the con-nection requests, which share the same trafc ow charac-teristics and involve signicantly lower capacities thanthose of the underlying wavelength channels, can be ef-ciently multiplexed or groomed onto the same wave-length/lightpath via simultaneous time and spaceswitching. Similarly, different trafc streams can bedemultiplexed from a single lambda-path.

3.2. Power requirements in network devices

The fundamental cause of energy consumption in elec-tronic equipment is the effect of loss during the transfer ofelectric charges, which in turn is caused by imperfectconductors and electrical isolators. Here, the consumptiondynamism, and hence are not easily applicable in a fullyadaptive online scenario or use power containment tech-niques based on switching off of inactive elements. In ourfully dynamic on-line approach, no switching off is as-sumed to be feasible (as explained in the previous section)and so, in this, it is completely different and not directlycomparable, in term of both performance and effective-ness, with all the previous ones.

3. Backgrounds

3.1. Wavelength routed networks

A wavelength-routed network, sometime also referredto as an optical circuit switched (OCS) network, is basicallycomposed of several OXC devices and opto-electronic edgerouters connected by a set of ber links. TheWDM technol-ogy is used to carve up the huge bandwidth available onthe optical bers into lower-capacity wavelengths (opticalchannels), which may be independently used to carryinformation across the same physical links. Circuitswitched connections, usually with high bandwidth andQoS-on-demand, are typically implemented by dynami-cally creating and tearing downmulti-hop optical channelsbetween client sub-networks according to a specic RWAstrategy. The above connections, called lightpaths, trans-parently traverse the ber network without being con-verted into an electrical signal. In some cases, they maypass through an optical/electrical/optical (O/E/O) conver-sion for regeneration, wavelength conversion or add/droppurposes. At the state of the art, there is still a large gapbetween the available capacity of an optical channel andthe much lower bandwidth requirements of a typical con-nection, but, on the other side, the number of wavelengthchannels (lambdas) available in most of the networks ofpractical size is much lower than the number of sourcedestination connections that need be made. Hence, trafcwhere a linear programming approach is proposed that isable to identify optimal topologies for given trafc loadsand generic network topology. In [23], various power-efcient grooming strategies, combined with lightpathextension and lightpath dropping, are evaluated in WDMnetworks where nodes have the tap-or-pass capability.

associated with a specic end-to-end circuit or the type

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2423rate depends on the transition frequency and the numberof gates involved, together with fabrication features (suchas architecture, degree of parallelism, and operatingvoltage). In pure transparent optical equipment, the mainenergy-hungry devices are the lasers, since the optical sig-nal has to reach the other end of the ber with sufcientquality in spite of the signal attenuation, dispersionand non-linear optical phenomena. Besides, the powerconsumption is also conditioned by sophisticated elec-tronic devices for coping with the technological complexityof the photonic environment. For example, when the in-volved ber strands need to cover long distances, severalintermediate electrical signal re-generators (3R) or opticalampliers (OA) are necessary (typically, an OA is neededevery 80100 km and a 3R every 5001000 km) to ensurethat the signal power and quality will be sufcient to reachthe other end of the ber with acceptable optical signal tonoise ratio (OSNR). Such OA and 3R have a not negligibleenergy cost that has to be taken into account when settingup the lightpath requests.

3.3. Energy-aware RWA

Introducing energy-awareness in RWA is based on theconcept of placing network trafc over a specic set ofpaths (sequences of nodes and communication links) sothat the overall network power demand and/or GHG emis-sions are minimized, while end-to-end connection require-ments are still satised. Typical infrastructures are denselymeshed, with many redundant interconnections amongnodes, so that many available paths can provide multiplereachability options between geographically distant sites.On such a mesh, wavelength routing is used to set-up alogical topology, which is then used at the IP layer for rout-ing. Every time a lightpath is established between any twonodes, the trafc of the lightpath will be handled as asingle IP hop by creating the abstraction of a virtual net-work topology on top of the physical one. This overlayapproach is based on the full separation of the routingfunctions at each layer, i.e. the connection routing/groom-ing at the IP layer is independent from the routing of wave-lengths at the optical layer. One of the key features of theabove model is rearrangeability, i.e. the ability to dynami-cally optimize the network as a consequence of the inde-pendence between the virtual and the physical topology.The above architectural exibility in building logical topol-ogies, together with physical connection redundancy andover-provisioning, provide fertile grounds for saving en-ergy, since a large number of available trafc routing anddevice management options can be exploited to optimizeenergy and carbon footprints network wide. Hence, en-ergy-aware logical network topologies, explicitly con-ceived to decrease power consumption in the operationalphase, can be dynamically built by minimizing the numberof energy-hungry devices traversed by the existing light-paths. In doing this, it is desirable to nd a good balancebetween the competing needs to avoid as many electricallypowered hops as possible (to reduce the power consump-tion at intermediate switching nodes, optical ampliersand regenerators) and avoid data transmission overexcessively long stretches, since moving data is quiteof energy source currently used by a network element[6]. Analogously, the same information has to be handledby control plane signaling protocols used for the reserva-tion and establishment of paths minimizing the use ofdirty power sources, as well as the overall energyconsumption, across the network.

4. The energy-aware network model

Dening a sustainable and effective network model tak-ing into account power consumption as well as energysource considerations is the essential prerequisite forintroducing energy-awareness within the wavelengthrouting context. A broad variety of NEs contribute to poweradsorption in a network: regenerators, ampliers, opto-electronic and totally optical routers and switches. Eachof these devices draws power in a specic way, which de-pends on their internal components and structures, on thetrafc load and on the relationship between the devices. Inaddition, some nodes may be powered by green energysources, while others may use traditional dirty energyplants; therefore, a differentiation between energy sourcesis required. NEs powered by green energy sources will notcontribute to the CO2 emissions but only to increase theoverall network energy consumption. In the dynamicRWA problem, the routing of connection requests is doneon a local optimality basis, i.e. considering the currentinformation available at the connection setup time. Thepotential of such an approach should not be underesti-mated; the conditions that determined such optimalitymay change, but the RWA strategies keeps its effectivenessas far as it is able to foresee the future network evolutions,both in terms of resources and energy utilization.energy-expensive. Energy consumption can be drasticallyreduced by maximizing the reuse of low-power transmis-sion links and highly connected devices, especially whenpowered by green sources, instead of obliviously spreadingthe trafc on the available routing/switching devices andcommunication resources. In other words, since a logicalnetwork topology is described by its constituent light-paths, a logical topology that minimizes the overall energyrequirement and the associated carbon footprint is one inwhich the choice of each individual lightpath, whilesatisfying the traditional RWA objectives and constraints,is driven by the above energy-efciency optimizationcriteria.

In order to support all the above behaviors, energy-related information associated with devices, interfacesand links need to be introduced as additional constraints(together with delay, bandwidth, physical impairments,etc.) in the formulations of dynamic RWA algorithms. Suchinformation must also be handled as new status features ineach network element that have to be considered in all therouting and trafc engineering decisions, and conveyed toall the various network devices within the same energy-management domain. This clearly requires modicationsto the current routing protocols by properly extendingthem to include energy-related information in their infor-mation exchange messages, such as the power demand

The above energy-related information and conceptsassociated with devices and links must be abstracted anddened in a formal and concise way into a comprehensivemodel that needs not to delve into unneeded details, butshould only describe the essential aspects needed to drivein an energy-conscious way the RWA algorithms andstrategies developed upon it. Therefore, we modeled thenetwork from a high-level perspective in an attempt tokeep the reference scenario as general as possible focusingon the effectiveness and energy-efciency of our ap-proach; the issues raised by modulation techniques,spectrum-sliced elastic networks, and other technologicalbreakthroughs, although interesting, fall outside the scopeof this paper, which is to provide an energy-awaredynamic RWA schema to route as many connections aspossible.

In detail, at the basis of our model we consider a mult-igraph G = (V,E) representing the network (Fig. 1), where Vis the set of nodes and E the set of edges, jVj = n, jEj =m. Nospecic assumption is made on the number of wavelengthsper ber link and on the number of bers on each link: anytwo nodes u, v 2 V may be connected by several edges(thus, multigraph). Each ber link (u,v) 2 E is characterizedby its physical length lu,v, together with the number ofavailable wavelengths wu,v. There can be more than one -

available on the physical circuit. Each tagged link (u,v)k,is characterized by its static global capacity au;vk and dy-namic residual capacity ru;vk . Clearly, for each link (u,v)k,its current load is given by au;vk ru;vk . Provided that asingle established lightpath or a chain of lightpaths be-tween the source and destination nodes has sufcientavailable capacity, each connection request can be routedonto that lightpath or chain. Also, a new lightpath maybe dynamically established, as the result of groomingdecisions.

The nodes of the graph model the routing and switchingdevices deployed in the network. We consider two types ofnodes: LERs (Lambda Edge Routers) and LSRs (LambdaSwitching Routers). LER nodes have both the electronicand optical interfaces, and have the capability to insert/extract traditional electronic trafc into/from the network.LSR nodes are OXC or recongurable optical add and dropmultiplexers (ROADMs) that are capable of switching thetrafc at wavelength level (since we model optical circuitswitched networks) and may be equipped or not withwavelength converters. Whenever an optical signal is con-verted into the electronic domain, it is implicitly assumedthat it is possible to apply 3R regeneration as well as wave-length conversion and add/drop at sub-wavelength granu-larity (grooming). Consequently, network trafc may be of

presen

2424 S. Ricciardi et al. / Computer Networks 56 (2012) 24202442ber connecting the same pair of nodes and, for simplicitysake, we assume that all the bers are of the same type(e.g. NZ-DSF ITU-T G.655/656), requiring an intermediateamplication or regeneration stage every K units of dis-tance. Typically K can assume the values KOA = 80 km fornative optical amplication systems and K3R = 5001000 km for 3R electric regeneration devices. On each berlink (u,v) there can be multiple wavelength links (u,v)k,modeled on the graph G as an additional virtual taggedlinks, where the tag k can be any of the wavelengths

Fig. 1. A network topology with two wavelengths per link (a) and its rerequest from node 1 to node 3) (c).two types: electronic time division-multiplexed (TDM)trafc (i.e. trafc that undergoes electronic processing)and pure optical trafc (i.e. WDM trafc entirely managedin the optical domain) with or without optical wavelengthconversion. Electronic routers have the ability to add/droptrafc into/from the network, to make electronic WC(Wavelength Conversion) and to regenerate the signal inthe electronic domain (3R regeneration). Optical routerssupport optical trafc with or without all-optical WC. Thatis, they may deect wavelengths through an electronic 3R

tations as multigraph (b) and as layered graph (assuming a connection

regenerator if the OSNR is too degraded and then switch

wanted uctuation in the green-biased node selection

WC and the latter consumes more power than optical

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2425function due to the temporary unavailability of the specicsource (e.g. sun, wind or tide), since the hosting sites areusually equipped with battery systems, ensuring the avail-ability of the accumulated green energy also during thesource off-times (e.g. the night hours for solar panels).Anyway, the devices powered by green energy should bealways preferred also for cost containment reasons, sincethe energy costs in the hosting sites/installations are regu-lated by signicantly advantageous contractual conditions.In fact most of the electricity providers and supplying util-ities apply some balancing policies for sites producing theirown energy from renewable source and placing the energyproduced in excess in the public electric grid (by reduc-ing the carbon footprint on a more global scale), so that,even in the case in which the required energy would becurrently extracted from a dirty source, its cost will be sig-nicantly lower when compared with other dirty-onlypowered sites.

4.1. Per-node power requirements

In order to characterize in a realistic and quantiableway the energy requirements of a specic network path(needed to accomplish our optimization goals within theRWA context), we need to estimate the power consump-tion of all the traversed NE (devices on the nodes andtransmission links) as a function of the involved trafctype. In doing this, we essentially consider two main trafctypes:

1. Electronic trafc, comprising add/drop, electronic WC,3R regeneration.

2. Optical trafc, with or without optical WC.

The above trafc types are characterized by a considerablydifferent power consumption when traversing a NE: elec-tronic trafc requires more power than optical trafc withthe wavelength through the corresponding output port[2]. In our network model, connections are bidirectionaland unsplittable, i.e. a trafc demand is routed over a sin-gle lightpath, and LER nodes can be source or destination ofa connection.

As for the energy, we derived a properly crafted per-node, per-link and per-lightpath energy model and powercost function, basing our estimation on the literature[13,14,15] and on the manufacturers technical sheets[24,25], with the aim of tting with the future energy-aware technologies that will adapt their power-consump-tion with their load [1014].

We distinguish between green and dirty energy sources,i.e. carbon-emitting and zero-carbon plants. Each noden 2 V has a statically associated attribute sn representingthe type of energy source (green or dirty) powering thecorresponding device, and we assume that green and dirtyenergy sources are heterogeneously distributed in thenetwork. This attribute has been kept static to avoid un-trafc without WC, due to the different devices involved[1,2].

Therefore, the power consumption of a specic light-path depends on:

1. The type of devices traversed along its route fromsource to destination node, e.g. router, switch, signalamplier/regenerator, etc.

2. The device class, in terms of hardware architectureand aggregated switching performance of the networkelement itself. More precisely, modular switching nodescapable to handle higher throughputs consume lessenergy per bit that smaller ones [26,27] since they aremore optimized and tend to be located in the centerof the network where the trafc is more aggregated,and opaque nodes equipped with electronic switch-ing matrices are more energy-hungry that their trans-parent photonic counterparts.

3. The type of trafc that it transports through each net-work element, i.e. electronic, optical with WC and opti-cal without WC.

The power consumption of real electronic and opticalswitching nodes with and without WC are reported in[1,2], where it can be observed that the electronic trafcgrows quickly with respect to the optical one and that,within the optical trafc context, the WC is the main factorinternal to the switching device requiring a not negligiblequantity of energy. In [14] it is shown that the base systemof an idle network device consumes approximately half ofthe total power drained by the device, while the other halfis consumed when the router is in its maximum congura-tion, i.e. maximum number of line cards/modules installedand operating at their full load. These power consumptionsrefer to commercially available devices whose architec-tures are not energy-aware: their power consumptionsonly slightly depend (23%) on the current trafc load,but strongly depend on the number of line cards installed[14,28]. Next-generation energy-aware routing/switchingnodes, designed with energy-efciency in mind and allow-ing dynamical adjustment of their power consumptionwith the variation of the trafc load by selectively puttinginto sleep or low-power mode some interfaces, line cards,and subsystems, will be characterized by a signicantlydominating load-dependent energy consumption compo-nent. However, by estimating the power demands usedin our model from the available quantitative data gatheredfrom the current devices implicitly forces the model tooperate in a worst-case situation making the achievedresults more comforting (since they will be greatlyimproved with the introduction of next generationenergy-aware devices). Therefore, even if actual routerarchitectures are not energy-aware, in the sense that theyconsume the same amount of power regardless of the traf-c load, here we consider future energy-aware architec-tures that scale their power consumption with theircurrent trafc load, thus giving rise to optimization[1,10,27].

Consequently to [14,28], we assume that the powerconsumption of a network element, modeled by a nodeor link, can be divided into two equal parts: xed andvariable power absorption. The xed power consump-tion is always present and is needed just for the deviceto stay on, while the variable power consumption de-pends on the actual trafc load that the device is currentlysupporting. The case in which the xed power consump-

trafc as the most power hungry devices. To this end, we

carefully balance the power consumption of all the possi-

In more formal terms, we dene for all the routing and

2/n lneBn

Bn x xBn ; 2

is the equation of the logarithmic function passing throughthe points (0,/) and (Bn,2/), modeling the best per-bit en-ergy consumption (i.e. optical trafc w/o WC in Fig. 3a)and:

and load (in Gbps) for different types of nodes (linear case).

of the Power consumption (y) as function of the load (x)assuming half xed (/) and half variable (m x)y = 1.5x + /y = 0.031x + /y = 0.01x + /

2426 S. Ricciardi et al. / Computer Networks 56 (2012) 24202442tion is signicantly greater or lower than the variable partwould affect more or less proportionally the optimizationgains. In particular, for the current energy-unaware de-vices, the xed part is much greater than the variable part(which only represents 23% of the total power consump-tion), leaving almost no space for optimization. In theopposite situation, if the xed power consumption wasmuch lower than the variable part, the likely outcomewould be that the optimization margins will increase alot; in this sense, our approach of assuming 50% for xedand 50% for variable power consumption can be consideredas conservative.

The previous considerations can be used to build a suf-ciently general per-node power consumption model.Starting from the power consumptions of the networkrouting devices (in Watts) as function of their aggregatedbandwidth (in Gbps) [1,2], we obtained the linear powerconsumption equations [10] reported in Table 1, whichwe used to calculate the real maximum power consump-tion of any kind of network node given its aggregatedbandwidth. In such a linear model, a slope ofmmeans thatfor each unit of trafc (Gbps) the router consumes m unitsof power (W). For example, an electronic router with anaggregated bandwidth of 10 Tbps is characterized by amaximum power absorption of 30 kW. An optical switchwith the same aggregated bandwidth consumes 0.62 kWwith WC and 0.2 kW without WC, which totally agree withthe values reported in [2,14].

Starting from such maximum power consumption val-ues, we obtain the curves in Fig. 2, in which the minimumpower consumption associated with the network device nin the idle state is given only by the xed power consump-tion /n of its base. The maximum power consumption 2/nis achieved when the node is fully loaded, i.e. when thecurrent load x is equal to the maximum aggregated band-width Bn of the node n. How the power consumption scalesbetween these two values has been studied carefully in[10]. In this work we observed that the power consump-tion associated with electronic trafc is higher than theone associated with optical trafc (Fig. 3a optical nodepower consumption not in scale; see the peak power con-sumption of optical nodes in Fig. 3b for in-scale values).Furthermore, we also observed that the power consump-tion of smaller nodes follows a worse trend with respect

Table 1Power consumption (in Watts) dependency laws on aggregated bandwidth

Node type Power consumption (y) as functionaggregated bandwidth (x)

Electronic y = 3xOptical w/ WC y = 0.062xOptical w/o WC y = 0.02xswitching nodes, a power consumption function Wn(x)expressing the power requirements of a node n character-ized by device-specic static consumption /n and perfor-mance class (aggregated bandwidth) Bn, variablyconditioned by a traversing trafc load x. The functionWn(x) can be viewed as a linear combination of the loga-rithmic function hn(x) and the line function #n(x) weightedby the parameter an(x):

Wnx anx hnx 1 anx #nx 1

where

hnx /n lne/n

Bn Bn x xBn

|{z}

variable power consumption

/n|{z}fixed powerconsumption

/n ble combinations between these two extremes.studied different power consumption functions, both pres-ent in literature and analytically conceived, to describe andto bigger ones, in which the per bit energy consumptionis lower (Fig. 3b). Therefore, we can delineate two ex-tremes: big nodes transporting optical trafc as the leastpower consumers, and small nodes transporting electronicFig. 2. Minimum and maximum power consumption of a network node.

of real routers from Table 1, taken in such a way to

Fig. 3. Power consumption functions of current, ideal and state-of-the-art for electronic and optical nodes (a) and for different sizes of nodes (b).

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2427#nx /nBn x|{z}variable powerconsumption

/n|{z}fixed powerconsumption

;

3

is the equation of the line function passing through thepoints (0,/) and (Bn,2/), modeling the worst per-bitenergy consumption (i.e. electronic trafc in Fig. 3a).

Bn is the capacity (aggregated bandwidth) of the routern (its performance class), and an(x) is the weighting param-eter between hn(x) and #n(x) depending on the class/perfor-mance of the router n and on the parameter b(x) whichaccounts for the specic type of trafc associated withthe load x that is actually passing through the node n:

Bnanx maxfBn;8n 2 Vg bx; 0 6 a 6 1; 4

Table 2Notation used in the energy model.

Parameter Energy model

/n Fixed power consumption of node nWn(x) Overall power consumption (xed + variable) of node n withhn(x) Logarithmic function#n(x) Line functionan(x) Linear combination weighting functione Eulers number (base of the natural logarithms)Bn Performance class of node n (aggregated bandwidth of all intb(x) Weighting function on the type of trafcm Slope of the power consumption functions (linear case)wu,v Number of wavelengths/channels crossing ber (u,v)Wu;vk x Power consumption of the link (u,v) on the wavelength k wilu,v Length of the ber (u,v) (km)guu;vk x Power consumption of the interface on the node u associatedQu,v Power consumption associated with the individual amplicaKOA Maximum allowed length of a link without need of optical aK3R Maximum allowed length of a link without need of 3R regenR(x) Power consumption associated with the individual 3R regeneWp(x) Power consumption of lightpath p with trafc load xlp Cumulative length of path p (km)penalize the more power consuming devices and trafctypes; e.g. for the electronic trafc, bx moptic w=o wcmelectronic 0:011:5 0:006.

2. an(x) weights one Eq. (2) or the other Eq. (3) functionaccording to the device class and trafc characteristicsbx 1 if the traffic x is optical w=o WC0:323 if the traffic x is optical w=WC0:006 if the traffic x is electronic

8>:

0 6 b 2 f1;0:323;0:006g 6 1: 5

Note that:

1. The values of b(x) have been obtained using the valuesof the involved NE.

trafc load x

erfaces)

th trafc load x

with the wavelength k on the ber (u,v) supporting the trafc load xtion device on link (u,v)mplication (km)eration (km)ration of the trafc load x (one for each wavelength)

bandwidth capacity) between two network nodes. In such

2428 S. Ricciardi et al. / Computer Networks 56 (2012) 242024423. The xed power consumptions /n of nodes are obtainedfrom [1,2,14].

For an explanation of the symbols used in the notationrefer to Table 2.

4.2. Per-link power requirements

End-to-end transmission links are characterized by apower consumption depending not only on the specic de-mand associated with the hardware interfaces located inboth the endpoints, but also on the impact introduced bythe optical amplication and regeneration devices neededby the signal to reach the endpoints with an acceptablequality, and thus, on the length of the traversed berstrands.

Accordingly, the power absorption of a transmissionlink realized on the wavelength k between the nodes uand v can be entirely described by the specic power de-mand characterizing the involved end-to-end interfacesplus the power required for powering all the possible inter-mediate regenerators or optical ampliers, if any. If theber between u and v is currently crossed by wu,v wave-lengths, we consider that the power requirements due tothe intermediate devices will be shared between the wu,vchannels simultaneously; typically, as for the OA, the en-tire frequency band (e.g. C-band) is amplied as a whole,without per wavelength granularity, whilst as for 3Rregeneration, per wavelength 3R is required. Thus thepower consumption Wu;vk x, associated with a link onthe wavelength k with load x traversing the ber (u,v),whose length is lu,v, can be described as:

Wu;vk x guu;vk x gvu;vk x

lu;vKOA

Qu;vwu;v

lu;vK3R

Rx;

6where

1. guu;vk x is the power consumption of the interface onthe node u associated with the wavelength k on theber (u,v) when operating at the minimum speedallowing to support a trafc load x without any loss ordelay increase. Here we do not take into account thevariable effect on power consumption of the dynamictrafc traversing the interface, whose impact is alreadyconsidered in the per-node consumption, and onlymodel the specic per-interface power demand accord-ing to its specic static hardware features (type of laser,its power, etc.) and to a multiple threshold scale charac-terizing its consumption depending on current operat-ing speed (implicitly dependent from the load x). Theabove guu;vk x function can be modeled as in [27].

2. Qu,v is the power consumption associated with the indi-vidual amplication device on link (u,v) (between 3 and15W) operating throughout the ber link (one for theentire frequency band); for simplicity, we assume thatall the amplication devices operating on the same linkhave the same power consumption.

3. R(x), dened in the sameway as #n(x) (i.e. electronic traf-c in a node), is the power consumption associated withthe individual 3R regeneration device of the trafc load xa dynamic scenario, connection requests have to be servedas soon as possible when they arrive; thus, we designedGreenSpark with simplicity in mind, which was consideredas a necessary requisite when developing the dynamicRWA scheme to keep as low as possible the computationalcomplexity. According to the typical assumptions in OCSnetworks, each connection is considered to be bidirectionaland consists of a specic set of trafc ows that cannot besplit between multiple paths. A connection can be routedon one or more (possibly chained) existing lightpaths be-tween the source and the destination nodes with sufcientavailable capacity or on a new lightpath dynamically builton the network upon the existing optical links. Connectionrouting and grooming decisions are taken instantaneouslyreecting an highly adaptive strategy that dynamicallytries to fulll the network resource utilization and connec-tion serviceability objectives together with minimizing theoverall power consumption by privileging cheaper (interms of power demands) chains of nodes/links and, be-tween them, trying to maximize the usage of devices pow-ered by green energy sources.

Without loss of generality, we route connection re-quests with only a constraint on the required bandwidth,and rely on the incorporation of other policies within thebandwidth-routing framework to perform routing basedon several QoS and impairment metrics such as limited la-tency, error rate, hop-count, delay, and losses. Such con-(one for each wavelength), when present; for simplicity,we assume that all the regenerationdevices operating onthe same link have the same power consumption.

4.3. Per-lightpath power requirements

Given a lightpath p as a sequence of nodes and taggedlinks (u,v)k with a trafc demand x, the power consump-tionWp(x) of p is given by the sum of the individual powerabsorption of all the traversed nodes and links plus the re-quired regenerations:

Wpx Pn2p

Wnx P

u;vk2pWu;vk x

lpK3R

Rx; 7

where lp is the cumulative path length. The third compo-nent in the Eq. (7) sum is needed to cope with fully trans-parent lightpaths whose length exceeds the maximumlength K3R that a signal can travel without need of 3Rregeneration. In this case, since all the intermediate de-vices do not convert the signal back and forth into electri-cal/optical form and only introduce impairments, a 3Rregeneration stage is required in correspondence with at

least lpK3R

j kintermediate nodes. If the lightpath p is fully

transparent (without wavelength conversion), the tag k isthe same on all the wavelength links.

5. The two-stage RWA scheme

The proposed energy-aware RWA scheme operates on-line, running at each request of a dedicated connectionwith specic service requirements (typically QoS on the

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2429straints can be incorporated into SLAs by converting theserequirements into a bandwidth requirement as shown in[29]. Impairments are accounted for by modeling 3R regen-eration into the framework, supported by the study in [30]in which impairment-awareness is included into theregeneration placement for WDM networks, and optoelec-tronic signal regeneration is employed to address the sig-nal quality of lightpaths that are found to be impairedwithout compromising the signal quality of any of thelightpaths.

The apparently conicting goals of minimizing cost andlength of designed paths while keeping the network re-source usage fairly balanced, and optimizing the overallpower consumption by reusing, as possible, energy-efcient paths across the network, give origin to a multi-variate and multi-objective optimization problem thatcan be solved according to a divide-et-impera strategy, setup of two-stages in which each stage separately handlesa specic objective by using properly crafted heuristics.Specically, in the rst stage (pre-selection phase), the goalis to determine an ordered list (whose length is dened bya parametric value k) of feasible minimum cost paths thatfully satisfy the connection demands, trying to leave oneach link of these paths sufcient room to satisfy furtherrequests as much as possible. Such strategy clearly impliesbalancing the load on all the available network resources.In the second stage (energy-aware decision phase), theproposed scheme analyzes, for each path found in the stageone, its power requirement as given by the aforementionedenergy model, and then selects the best available solutionaccording to several heuristic criteria based on nding agood compromise between the traditional carriers objec-tives and the new green requirements (i.e. limiting powerconsumption or using green energy sources to reduceGHG emissions).

Towards this goal, we explicitly dened and studied twodifferent green optimization objectives: the rst one, aim-ing at reducing the power consumption throughout thenetwork, and hence its operating expenditures; the secondone, oriented to minimize the network carbon footprint onthe environment by minimizing the GHG emissions.

5.1. Prerequisite control plane facilities

The proposed scheme also requires several forms ofcooperation between the nodes concurring to the RWAproblem solution. This implies that every node needs torun distributed control-plane services (such as those pro-vided by the GMPLS framework) keeping up-to-date infor-mation about the complete network topology, resourceusage and power demand attributes, as well as taking careof resource reservation, allocation, and release.

More precisely, a periodic link-state advertisement(LSA) schememust convey all the link and node state infor-mation (including energy related ones) to every node inthe network, ensuring the complete synchronization be-tween all the nodes network status views. Since theamount of per-link state information is very small, anyappropriate enhanced link state scheme like those em-ployed by OSPF can be adequate for this purpose, like theone developed in [6].The Dijkstra-based path selection scheme of stage oneshould meet certain conditions:

1. A link may not reserve more trafc than it has capacityfor.

2. Shorter paths should be preferred when they consumefewer network and energy resources.

3. Critical resources, e.g. residual bandwidth in bottlenecklinks, should be preserved for future demands.

The last two conditions reect that what we really seek isto keep the connection blocking probability (or, in otherwords, the rejection ratio) as low as possible, or equiva-lently to increase as much as possible the networkutilization.

In addition, an extended signaling/reservation protocol,such as RSVP-TE, can be used to setup and release pathsand lightpaths and handle all the bandwidth, ber orwavelength resources reservation and allocation/dealloca-tion operations required during such activities. In detail, asa new request arrives, the control plane on each node,starting from the originating one, runs our source-basedlocalized RWA algorithm, calculates the new overlay net-work topology and triggers the proper path setup actionsby sending a reservation request toward the destinationand provisionally reserving bandwidth resources. TheRWA scheme, operating according to a two-layer model(i.e. an underlying pure optical wavelength routed networkcore and an opto-electronic time division multiplexedlayer built over it) should determine if the request can berouted on one of the already available lightpaths, bytime-division multiplexing it together with other alreadyestablished connections, or a new lightpath is needed onthe optical transport core to join the terminating (edge)nodes. In presence of multiple options between new feasi-ble and already established lightpaths, the link weightingand path selection functions of the two stages, applied onthe existing lightpaths and to the wavelength links thatcan be used to set up new lightpaths, together with the en-ergy costs, dynamically determine the best compromise(between network and energy costs) routing solution forthe request, starting from the current network status. Forexample, if two lightpaths between source and destinationexist, both with sufcient available capacity, if the differ-ence in network cost between them falls below a specicacceptability threshold, the tie is resolved in favor of thegreener lightpath. Such policy guarantees maximum light-path utilization and automatically achieves, as long as pos-sible, effective dynamic grooming and power usage,assuming that the topology (link state) database is prop-erly updated.

The signaling scheme for triggering the new lightpathset-up and reserving the required bandwidth, ber orwavelength resources along the path is very similar tothe RSVP-TE protocol used by GMPLS. To make a reserva-tion request, the source node needs the path and the band-width that it is trying to reserve. The request is sent by thesource along with path information. At every hop, the nodedetermines if adequate bandwidth is available in the on-ward link. If the available bandwidth is inadequate, thenode rejects the requests and sends a response back to

ler (low global capacity au;vk ) and more congested (low

2430 S. Ricciardi et al. / Computer Networks 56 (2012) 24202442residual capacity ru;vk ) links. Note that every two linkswith the same residual/maximum capacity ratio but differ-ent residual or maximum capacity values will have differ-ent associated weights. This avoids assigning the sameweight to two links with the same saturation ratio but withdifferent residual or global capacity. Therefore, we choosethe weighting function (8), which satises all the desiredproperties discussed before. Note also that the rst stageis exclusively based on load-balancing criteria, and no en-ergy consideration is present at all; this guarantees thatthe k paths selected in this stage are the best balancedones, thus giving priority to the traditional network opti-mization criteria of minimizing the connections blockingthe source. If the bandwidth is available, it is provisionallyreserved, and the request packet is forwarded onto thenext hop in the path. If the request packet successfullyreaches the destination, the destination acknowledges itby sending a reservation packet back along the same path.As each node in the path sees the reservation packet, itconrms the provisional reservation of bandwidth. In addi-tion, it also performs the required conguration needed tosupport the incoming trafc such as setting up labels in aGMPLS label switching node, or reconguring the lambdaswitching internal devices (such as MEMS) in a transparentoptical wavelength switching system.

5.2. The rst stage: selecting the candidate paths

The rst stage of the GreenSpark RWA schema com-putes a list of k feasible cycle-free paths, in increasing or-der of cost, between the source and the destinationnodes of the connection to be routed, constrained by itsQoS requirements. Here k is a congurable parameter thatcan be used to limit the number of feasible paths thatshould be considered in the following step, thus control-ling the depth and granularity of the analysis processaccording to a performance/precision compromise. TheK-SPF (k-shortest paths rst) algorithm used has beenexplicitly modied to meet the specied bandwidthrequirements of each new request and to enforce thewavelength continuity constraint so that, when traversingconverter nodes, we are totally free in selecting any outgo-ing link of the multigraph (i.e. any wavelength), whereaswith all the other nodes we can only select an outgoinglink corresponding to the same wavelength associatedwith the incoming one. The above pre-selection processis driven by a link weighting function x((u,v)k) taking intoaccount, for each link (u,v)k, the (static) global capacityau;vk and the current (dynamic) residual capacity ru;vk stillavailable on the link. Intuitively, a good weighting functionshould be inversely proportional to both the residual andthe maximum capacities, but the contribution of thesetwo factors need not be the same. Following the analysisin [8], the link weighting function is dened as:

x : E ! R;xu;vk ru;vk logau;vk1 8Such a function exhibits the desirable property of lead-

ing to a good load-balancing over the network, since it triesto avoid bottleneck link by assigning higher costs to smal-ratio. Energy-awareness is introduced only in the secondstage, where the greenest path among the k best-balancedcandidate paths is nally selected.

5.3. Second stage: choosing the best path

The k minimum cost paths found by the K-SPF algo-rithm in the rst stage are the k best paths as for networksblocking probability (the percentage of rejected connectionrequests), since the weighting function x((u,v)k) tends tobalance as much as possible the use of the network re-sources. Among these k best-balanced paths, we now haveto choose the optimal path among them according to ourenergy-aware selection criteria, aiming at minimizing thepower consumption or the carbon footprint. For this pur-pose we need to introduce a properly crafted heuristicworking as a path scoring function, to differentiate amongthe available preselected paths and choose the most en-ergy-efcient one. The scoring function fS is dened onthe set of all the possible paths P and will evaluate thepower consumption and carbon footprint of the k pathsK = {pi, i = 1,2, . . . ,k} obtained from the rst step:

fSp : P! R: 9The (total) power consumption Wp(x) of a path p de-

ned in eq. (7) can be decomposed as the sum of the powerconsumption of the traversed devices that are powered bygreen WGpx and dirty WDpx energy sources:

Wpx WGpx WDpx: 10Note that the carbon footprint of a lightpath is only gi-

ven by the power consumption of the involved NEs that arepowered by dirty energy sources, as the NEs powered bygreen energy sources do not contribute to GHG emissions.

Therefore, if our primary objective is to minimize theGHG emissions (GreenSpark MinGas), we have to choosethe path p which has the lowest carbon footprint WDpx(primary objective) and, among paths with the same min-imum carbon footprint (if any), we choose the path thatminimizes the total power consumption Wp(x) (secondaryobjective):

fSp WDpx logWpx: 11Analogously, if our main goal is reducing the overall

power consumption and, thus, the network operating en-ergy costs (GreenSpark MinPower), we need to use anobjective function privileging the paths with minimal totalpower consumptionWp(x) and, among them, choosing theone with the minimum carbon footprint WDpx:

fSp Wpx logWDpx: 12The computation of the scoring function is done for

each of the kminimum cost paths, and the path p eventu-ally chosen is the one with the lowest fS(p) value:

p argminffSpjp 2 Kg: 13If more than one such lightpaths exist (i.e. with the

lowest fS (p) value), the one with the minimum i indexin the set of lightpaths K is selected (to maximizeload-balancing).

The path p is the best path between the best load-balanced paths that minimizes the carbon footprint oroverall power consumption according to the proposed en-ergy model, and it will be used to route the connectionrequest.

Note that the rst stage cost function is dened over theset of edges (8), whereas the energy-aware scoring func-tion is dened over paths (9) to reect our intent of achiev-ing an acceptable compromise between the traditionalnetwork optimization objectives, typically based only onspecic link properties, and the energy-related ones thatneed to take into account more complex considerationsto be done on the higher layer concepts of lightpath/chan-nel, interface and node role and wavelength processingpractice such as optical amplication and 3R regeneration.That is why we structure the decision process into twoindependent phases and select among the k candidatepaths the one with the minimum carbon footprint orpower consumption according respectively to the func-tions (11) and (12), instead of simply selecting the mini-mum cost path based only on the traditional costfunction (8).

The generic GreenSpark algorithm is sketched in Fig. 4.

6. Time and space complexity analysis

In the rst stage, the computing of the K-SPF for ndingthe k feasible paths for a specied sourcedestinationpair requires in the worst case a time complexity ofO(k (m + n logn)) [31]. The second stage computes theobjective function fS for each of the k paths found. The func-tion calculation requires the computation of the power con-sumption or GHG emissions for each network element inthe path. The maximum length of a cycle-free path in agraph with n nodes is n 1, thus the second stage requiresO(k n). Hence, since the K-SPF complexity is thedominating factor between the two stages, the worst caseruntime is given by the polynomial time complexityO(k (m + n logn)). Therefore, GreenSpark belongs to thesame polynomial complexity class of the fastest SPF im-proved by using a priority queue with a Fibonacci heap inthe implementation, O(m + n logn) [32]. GreenSpark com-plexity is also lower than the quadratic complexity of theoriginal SPF algorithm, O(n2), and signicantly lower thanthe cubic complexity of nave MIRA O(n3m log (n2/m))optimized with the Goldberg max-ow algorithm [33].

As for space complexity, our multigraph network repre-

enSpa

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2431The algorithm takes as input the current network state G,the connection request q = (s,d,b) between node s and dwith QoS bandwidth requirement b, the k parameter ofthe K-SPF and the objective function fS. In the rst stage(lines 12), the K-SPF algorithm nds the k minimum costpaths with sufcient free bandwidth connecting the sourceand destination nodes. The k paths are the best onesaccording to the load-balancing cost function x((u,v)k) ofEq. (8). In the second stage (lines 37), the chosen objectivefunction fS is evaluated for the k minimum cost paths andthe best path p is eventually chosen to route the connec-tion request q. Finally, the new x((u,v)k) costs are updatedonly for the edges of the path p, and the chosen path andnew network state are returned.

Fig. 4. The Gresentation requires less space with respect to the layeredgraph approach conventionally used in dynamic RWA algo-rithms (Fig. 1). Using up to kwavelengths on each edge, thelayered representation with C converter nodes will requirekn + 2 nodes (k layers, each dedicated to an individualwavelength, plus two additional nodes to serve as ingressand egress) and km + 2k + C (k 1) edges (converters canbe modeled by cross-layer edges that connect each layerto the k adjacent layer a wavelength conversion spanningmultiple frequencies will thus entail many such edges insequence), whilst the equivalent multigraph representa-tion will require only n nodes and km edges, thus notablyreducing the space complexity. Besides, in the layeredgraph, the ingress and egress nodes as well as the edges

rk algorithm.

connecting them to the network have to be built each timea new connection arrives, whilst in the multigraph ap-proach this preprocessing phase is not necessary thanksto its compact representation. Note that, even in absenceof wavelength conversion, all the layers of the layeredgraph have to be explored, since the (rst) wavelength ofthe lightpath may be any, which compensates the addi-tional check needed in the multigraph approach to enforcethe wavelength continuity constraint. Furthermore, thehigher number of nodes and edges required by the layeredgraph with respect to the multigraph approach increasesthe time complexity which strictly depends on the n andm parameters.

The low computational and space complexity requiredby the GreenSpark algorithm with the multigraph networkrepresentation helps lightening the computational burdenof path computing elements and serving the connectionswith lower delay with respect to more complexapproaches.

7. Performance evaluation and results analysis

In order to evaluate the effectiveness of the GreenSparkenergy-aware RWA framework and its impact on thepower consumption and carbon footprint of telecommuni-cation networks, we conducted an extensive simulation

study on the network topology modeled as undirectedgraphs in which each link has a non-negative capacityand a specic power demand depending on both its phys-ical and technological features. All the nodes in the graphare characterized, apart from the traditional network-levelcapabilities such as wavelength conversion and add-and-drop capability, by their power absorption and type of en-ergy source (i.e. green or dirty), as dened in the energymodel of Section 4.

To improve the signicance of the obtained results andmake them more easily comparable with the other experi-ences available in literature, we spent a signicant efforton the use of realistic data in all our experiments (networktopology, trafc demands, costs, and power consumptionmodels). Accordingly, we used in our simulations thewell-known network topology Geant2 [34] of Fig. 5 withthe bandwidths for the links ranging from OC-1 to OC-768 bandwidth units. Here, trafc demands have beenmodeled by using different randomly generated or staticpredened [35,36] trafc matrices. In the latter case, thetrafc volumes have been scaled proportionally to the re-ported trafc distributions. The energy model has beenfed with the realistic power consumption values associatedwith nodes and links taken from [2,10,37]. Recall that, sinceno per-node sleep mode is assumed to be possible, thenetwork elements are always powered on and thereforethe GreenSpark algorithm bases its decisions exclusively

rk top

2432 S. Ricciardi et al. / Computer Networks 56 (2012) 24202442Fig. 5. Geant2: real netwo ology used in simulations.

on the variable power consumption part, which is the onlyone that can vary and thus be optimized. Power consump-tion results refer thus only to the variable powerconsumption.

Each connection request was characterized by a band-width demand ranging from OC-1 to OC-192 units (i.e.from 50 Mbps up to 10 Gbps). As the network load grows,that is, the number of busy connection resources increasesmore and more with respect to the free/released ones, wecontinuously monitored the overall network power de-mand, the percentage of green energy used compared tothe maximum available, and the network efciency ex-

minimum interference routing algorithm (MIRA) [40], al-ready implemented in several commercial solutions, givesus a real portrait of the power and GHG savings that will beconsequent to the introduction of the proposed schemawithin real world infrastructures, and, at the same time,demonstrates the absence of signicant performance bur-dens in traditional network management objectives(increasing blocking probability, reduced load-balancing,etc.) due to the new energy optimization goals.

The behavior of the algorithm varying the k parameterhas been extensively studied (Section 7.2) and, for the con-sidered network topology, an optimal value of k = 3 was

probability. However, starting from 500 connection re-

e Gea

ing fr, 12, 25

m, 100ing ac, MIRked co

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2433pressed by the rejection ratio/blocking factor. All the sim-ulation experience has been conducted in a properlycrafted optical network simulation environment [38] thatallows the creation of network topologies along with thespecication of simulation parameters and congurationles. All the results have been determined with a 95% con-dence interval not exceeding 6% of the indicated values,estimated by using the batch means method with at least25 batches. All the runs have been performed on an Intel

Core i7-950 CPU @ 3.07 GHz with 16 GB RAM and 64 bitoperating system server running Sun Java Runtime Envi-ronment v.1.6. In all the experiments, we used a dynamictrafc model in which connection requests, dened by aPoisson process, arrive with a parametric rate of c re-quests/s and the session-holding time is exponentially dis-tributed. The connections are distributed on the availablenetwork nodes according to the above random-generatedor predened trafc matrices, as summarized in Table 3.

In our lambda-switched optical framework, the re-sources occupied by the routed connections are countedas the sum of the ratio between the free and the busybandwidths along the edges. Resources are thus repre-sented as the sum of the bandwidths on all the networkedges, while the trafc volume is represented by the quan-tity of the utilized bandwidth in a certain time.

In all the experiments, GreenSpark has not been com-pared with other analogous power-containment solutionsknown in literature because, at the state-of-the-art, almostall the available schemes achieve their savings by power-ing off interfaces or entire nodes (practice avoided in realnetwork as already mentioned in the introduction), so thatthe comparison would be misleading since shutting downan entire device would cancel its xed power consumptionwhich, in our always-on approach, is present all the time.Conversely, the use of provably efcient and publicly avail-able algorithms such as min hop algorithm (MHA) [39] and

Table 3Parameters used in the simulations.

Simulation parameters Dant

Number of connections VaryRandom generated bandwidths {1, 3GreenSpark k 1, 3,Spark k, kHop 3, 20KOA,K3R 80 kSource, destination VaryRWA algorithms MHAMeasurements Blocquests, its blocking probability grows at quite a fast pace.All the algorithms belonging to the Spark family (Spark,GreenSpark MinPower and GreenSparkMinGas) performsensibly better than the other ones, and all their versionsachieve similar and very satisfactory results in terms ofthe connection blocking probability.

In Fig. 7 we compared the total power consumption(green and dirty) obtained by the different algorithms ver-sus the connection requests. MIRA reveals to be the highestpower consumer, followed by MHA which, in contrast with

nt2 network

om 0 to 3000 with different resolutions4, 48, 192} OC-units with different distribution probability

0 kmcording to a Poisson process, duration times exponentially distributedA, Spark, GreenSpark MinPower, GreenSpark MinGasnnections, power consumptions, green energy percentageschosen as the best compromise between the different opti-mization objectives of the two stages (load-balancing andgreenness) and time performance (recall from Section 6that the complexity depends on k). However, for clearnesssake, we rst show the results of the comparative simula-tions with the other RWA algorithms (Section 7.1) and,then, show how the biasing of the k parameter affectsthe performance (in terms of the two stage objective)and the time complexity of GreenSpark for the given net-work topology.

7.1. Comparative simulation study

In the rst set of simulation shows, we report the com-parison of the GreenSpark framework with other well-known RWA algorithms. In these tests, the k parameterof GreenSpark has been tuned to an optimal value (k = 3)for the considered network topology, as a result of theextensive simulation study reported in Section 7.2.

In Fig. 6 we plotted the connection blocking probabilityversus the generated connection requests. We can observehow MHA exhibits the highest blocking probability, essen-tially due to the congestion of the communication linksassociated with the shortest paths. MIRA [40] improvesthe performance of MHA, and achieves lower blocking

the previous graph, performs better than MIRA. This is due the connections (and, thus, statistically introduces less

Fig. 6. Connection blocking probability versus connection requests.

2434 S. Ricciardi et al. / Computer Networks 56 (2012) 24202442to the longer paths chosen by MIRA with respect to MHAthat, in turn, always chooses the shortest paths to routeFig. 7. Total power versuspower consumption). In this graphic, we can also observethe rst big difference inside the Spark family: theconnection requests.

GreenSpark algorithms have lower total power consump-tion with respect to the energy-unaware Spark. Besides,we note that the two GreenSpark algorithms, MinPower

and MinGas, perform almost the same as for the totalpower consumption, with MinPower doing slightly better,as expected. Anyway, the fact that the two GreenSpark

Fig. 8. Green power (percentage) versus load (routed connections).

ection

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2435Fig. 9. Load (routed conn s) versus power budget.

algorithms have almost the same total power consumptiondoes not mean that their carbon footprint is the same. Thisleads us to the next graphic of Fig. 8, in which the greencomponent of the power consumption has been reported.

The results, highlight that there is big difference in theuse of green energy depending on the chosen GreenSparkoptimization goal, as well as for the other algorithms.GreenSpark MinGas exhibits the topmost green powerusage percentage, i.e. it prefers lightpaths passing throughgreen-powered NEs and avoids sites powered by dirtysources as much as possible. More than 29% of the totalpower used by GreenSpark MinGas comes from green en-ergy sources, thus saving considerable quantity of CO2from being emitted by the network during its operations.The other two algorithms of the Spark family are charac-terized by a lower green power usage, as expected, whilekeeping the blocking probability unaffected. Althoughbeing energy-unaware, MIRA performed quite well in ourtests in terms of green power usage percentage, basicallydue to its minimum interference driving criteria whichtends to balance the usage of network resources (whoseenergy is equally distributed among green and dirty energysources), even if it exhibits a very high connection blockingprobability, reaching values of 54% starting from a load ofjust 1450 connections. MHA exhibits the worst perfor-mance both for the green power usage and for the connec-tion blocking probability, showing its limitations incomplex network scenarios where a number of constraints,comprising the energy-efciency ones, have to be takeninto account. A particularly interesting issue comes from

the observation of the pseudo-sinusoidal trend in the useof the green resources characterizing all the algorithms ofthe Spark family. This behavior is due to the specic costand scoring functions associated with this family, in whichless costly/greener paths will be chosen rst, making thegreen energy percentage rise. As the usage of green pathsraises, however, also the dynamic cost assigned to suchpaths increases as a consequence of their increased load(according to the load-balancing criteria of Eq. (8), untilalternative non-green paths will be cheaper than the greenones and thus will be preferred for connections routing.This will make the green energy percentage decrease,but, at the same time, increase the cost of these alternativepaths, until it will be again more convenient to route theincoming connections on green paths, and so on. Therefore,the pseudo-sinusoidal trend of the Spark family is some-how a visual proof of the efciency of the two phase selec-tion scheme which, at rst, tries to balance the networkload and, then, to minimize the specic scoring function,such as the total power (GreenSpark MinPower), the totalGHG emissions (GreenSpark MinGas) or the total cost(Spark).

Starting from the consideration that it is a commonpractice that network operators contract a xed powerbudget with their energy supplier and then strive to re-main within that budget since surpassing the thresholdwill result in high penalty rates on the overall energy costs,in Fig. 9 we plotted the load versus the power budget re-quired to route the connections. The energy-aware Green-Spark algorithms exhibit an almost optimal growth trend,

us dif

2436 S. Ricciardi et al. / Computer Networks 56 (2012) 24202442Fig. 10. Total power (green + dirty) and load (routed connections) vers ferent algorithms at a load causing a blocking probability of 0.05 (5%).

showing the highest increase in the load against a xedincrease in the power budget with respect to the otheralgorithms. We can observe that the entire set of connec-tions can be routed in the network keeping its power

budget below the 70 kW threshold (and the blockingprobability at the lowest observed values). From this pointof view, Spark performs notably well, considering that it isenergy-unaware: its power budget is only 83 kW. It is

Fig. 11. Connection blocking probability versus connection requests.

ersus

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2437Fig. 12. Total power v connection requests.

worthwhile to note that inside the power budget of theSpark algorithms, there are much many connections (more

than 2000) with respect to the MHA and MIRA algorithms,which only route between 1200 and 1400 connections.

Fig. 13. Green power (percentage) versus load (routed connections).

nectio

2438 S. Ricciardi et al. / Computer Networks 56 (2012) 24202442Fig. 14. Load (routed con ns) versus power budget.

Furthermore, their power budgets are sensibly higher,being 140 kW and 195 kW respectively, between two andthree times more than GreenSpark.

The last test of this series was conducted by keeping theQoS on the blocking probability at the constant value of 5%.We measured both the total required power and the loadoffered by the algorithms. The graphic in Fig. 10 clearlyshows the great improvements introduced by the Sparkfamily both for the power consumption and the routedconnections. In particular, we observe that the GreenSparkalgorithms need as low as half of the power required byMIRA and MHA and route double the number of theirconnections, making them very attractive also for QoS con-strained networks with strict requirements on the connec-tion blocking probability.

As a conclusion, we observe that there is a generationgap between the Spark family and the traditional MIRAand MHA algorithms, both in terms of power consumptionand blocking probability. In particular, Spark performancesare quite satisfactory, but GreenSpark algorithms, thanksto the two stages load-balancing and green objectives bal-anced by the k parameter, perform much better in terms ofpower and GHG, with MinGas even superior than MinPow-er, since it considerably lowers the GHG emissions whilekeeping almost the same total power requirements thanMinPower. Results showed that GreenSpark algorithmsnot only signicantly lower the required power and GHGemissions but also increase the connections acceptance

ratio, showing that properly crafted RWA algorithms canenable greener networks with even better performancethan before.

7.2. Tuning the GreenSpark k parameter

The k parameter value biases the load-balancing criteriaof stage one (Eq. (8)) and the greenness criteria of stagetwo (Eq. (11) and Eq. (12)). Stage one restricts the set ofpossible paths to the best balanced k paths between in-gress and egress nodes according to its cost functionx((u,v)k); stage two selects, among such paths, the green-est one according to its scoring function (MinGas orMinPower). At the extreme cases, a k value of 1 wouldrestrict the stage one to always select the minimum costpath (thus, the best load-balanced path according toEq. (8)) and the stage two to always select the only pathavailable from stage one, making the algorithm totallyenergy-unaware, and therefore reducing it to a simpleDijkstra-based weighted shortest path (that is a minimumcost one); from the other side, a large enough k value(greater than the maximum number of possible paths be-tween any two nodes) would make the algorithm totallygreen, completely discarding the load-balancing effectof stage one. Therefore, smaller values of the k parameterbias the solution by privileging well balanced paths, whilelarger values of the k parameter privilege energy relatedobjectives rather than the traditional network

eenSp

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2439Fig. 15. Total power (green + dirty) & Load (routed connections) versus Gr(5%).ark with varying k values at a load causing a blocking probability of 0.05

8. Conclusions and future work

References

2440 S. Ricciardi et al. / Computer Networks 56 (2012) 24202442management ones. A in depth study of the k parameter istherefore interesting and it is reported here for the Geant2network topology.

In Fig. 11 we show the blocking probability of Green-Spark with varying values of the k parameter versus theconnection requests. As expected, the k value affects verylittle the blocking ratio since the actual load-balancing isdone in the rst stage, whose output are always the k bestbalanced paths; in this sense, load-balancing is assured bystage one.

On the contrary, if we look at the total (green + dirty)power consumption versus the connection requests re-ported in Fig. 12, we observe that the higher the k value,the lower the total power consumption, as expected. Infact, with higher k values, the second stage will have thepossibility to choose among a greater number of alterna-tive paths to minimize ecological footprint of the network,both for MinPower and MinGas, which indeed performvery similar; in this sense, the green aspect is committedto stage two.

However, if we take a look inside the total power con-sumption at the green power percentage versus the loadreported in Fig. 13, we see that there is a notable differencein the use of green energy sources at high k values. Withk = 5 and k = 7, the green power percentage of GreenSparkMinGas markedly increases (whilst the MinPower is in theaverage as expected being GHG-unaware), showing thatthere is still room for green optimization at the expenseof additional computational complexity due to the calcula-tion of the higher number of alternative paths in the stageone. This result suggested the basis of a future work ofours, in which we will try to reach such greener pathsthrough a one-stage algorithm with an omni-comprehen-sive energy-aware/load-balancing cost function employedto directly achieve such paths. We also note that the greenpower percentage for k = 1 is quite high, showing that theload-balancing may have positive effects on the GHG emis-sions when the energy sources are heterogeneously dis-tributed in the network.

In Fig. 14 we plotted the load versus the required powerbudget for different k values. As seen in Fig. 12, with thesame k value, MinPower and MinGas perform quite simi-larly in terms of total power consumption, but GreenSparkwill require different power budgets depending on the kvalue. The higher the k, the lower the power budget butalso the higher the computational complexity requiredfor the path calculation at each connection request set-uptime. Anyway, it is worthwhile to note that there is greatimprovement between k = 1 and k = 3, and only limitedgain for greater values, meaning that already with k = 3alternative paths the GreenSpark framework is able tosensibly reduce the power budget in an optimal balancebetween greenness and performance.

Finally, the results of the test on the QoS on the blockingprobability at the constant value of 5% is shown in Fig. 15.The total power required decreases with the increase of thek parameter value but again, while from k = 1 to k = 3 thereis a great reduction of the total power, when passing fromk = 3 to k = 5 and k = 7 there is no such a great benet. Notealso that, at such low load (5% of blocking probability),there is no great difference in varying the k values as for[1] S. Aleksic, Analysis of power consumption in future high-capacitynetwork nodes, IEEE/OSA Journal of Optical Communications andNetworking 1 (3) (2009) 245258, http://dx.doi.org/10.1364/JOCN.1.000245.

[2] B. Project, WP 21 TP green optical networks, D21.2b report on Y1 andupdated plan for activities, 2009.In this work, we focused our research effort on Green-Spark, a novel heuristic-driven dynamic RWA frameworkaiming at the minimization of power consumption andGHG emissions in wavelength routed backbone networks.GreenSpark operates by progressively routing the dynami-cally incoming connections on a two-stage basis; in therst stage, a set of k feasible paths is found according totraditional load-balancing objective. Then, in the secondstage, the greenness of the k paths is evaluated both interms of power consumption (MinPower) and GHG emis-sions (MinGas), and the greenest path is nally selectedto route the connection. Even with low k values (i.e.k = 3), and despite its very low computational complexity,GreenSpark achieves signicant power savings and carbonfootprint reduction together with an increment of theload-balance, resulting in lower blocking probability ascompared with several widely used routing algorithms,as veried by the extensive simulation study.

Apart from dening an energy consumption model forthe IP over WDM network, one of the most signicantadded values of the framework is the incorporation of bothphysical layer issues, such as power demand of each com-ponent, and virtual topology-based energy managementwith integrated trafc grooming, adversely conditioningthe usage of energy hungry links and devices. Moreover,since the above model also takes into account the type ofpower supply associated with each device, by privileginggreen sources, the proposed scheme can also be usefulfor equalizing the carbon footprint of entire areas withina real network scenario in which each device locationmay be characterized by a differentiated (green or dirty)energy source. Here, multi-objective optimization mayhelp us in nding the appropriate trade-off according tothe relative importance of network performance and envi-ronmental friendliness.

As future work, we are studying an omni-comprehen-sive energy-aware/load-balancing cost function to directlynd green paths in a single stage with even lower compu-tational complexity. In addition, we are investigating newenergy-aware trafc engineering strategies and networkre-optimization methods, aiming at dynamically reducingpower demand, GHG emissions and costs on a time basis,by moving data wherever electricity costs are lowest at aparticular time.the number routed connections since most of the connec-tions will have sufcient resources to be routed, even if aslightly better performance is observed in correspondenceof the k = 3, i.e. when both the load-balancing and thegreenness objectives are fairly weighted.

S. Ricciardi et al. / Computer Networks 56 (2012) 24202442 2441[3] J. Baliga, R. Ayre, K. Hinton, W. Sorin, R. Tucker, Energy consumptionin optical IP networks, Journal of Lightwave Technology 27 (13)(2009) 23912403, http://dx.doi.org/10.1109/JLT.2008.2010142.

[4] G. Shen, R. Tucker, Energy-minimized design for IP overWDM networks,IEEE/OSA Journal of Optical Communications and Networking 1 (1)(2009) 176186, http://dx.doi.org/10.1364/JOCN.1.000176.

[5] F. Idzikowski, S. Orlowski, C. Raack, H. Woesner, A. Wolisz, Savingenergy in IP-over-WDM networks by switching off line cards in low-demand scenarios, in: OpticalNetworkDesign andModeling (ONDM),2010 14th Conference on, 2010, pp. 16. doi:http://dx.doi.org/10.1109/ONDM.2010.5431569.

[6] J. Wang, S. Ruepp, A.V. Manolova, L. Dittmann, S. Ricciardi, D.Careglio, Green-aware routing in GMPLS networks, in: Computing,Networking and Communications (ICNC), 2012 InternationalConference on, 2012, pp. 227231. http://dx.doi.org/10.1109/ICCNC.2012.6167416.

[7] I. Chlamtac, A. Ganz, G. Karmi, Lightpath communications: anapproach to high bandwidth optical wans, IEEE Transactions onCommunications 40 (7) (1992) 11711182, http://dx.doi.org/10.1109/26.153361.

[8] F. Palmieri, U. Fiore, S. Ricciardi, SPARK: a smart parametric onlineRWA algorithm, Journal of Communications and Networks 9 (4)(2007) 368376.

[9] S. Nedevschi, L. Popa, G. Iannaccone, S. Ratnasamy, D. Wetherall,Reducing network energy consumption via sleeping and rate-adaptation, in: Proceedings of the 5th USENIX Symposium onNetworked Systems Design and Implementation, NSDI08, 2008,pp. 323336.

[10] S. Ricciardi, D. Careglio, G. Santos-Boada, J. Sol-Pareta, U. Fiore, F.Palmieri, Towards an energy-aware internet: modeling a cross-layeroptimization approach, Telecommunication Systems (2011) 2122.

[11] A. Muhammad, P. Monti, I. Cerutti, L. Wosinska, P. Castoldi, A.Tzanakaki, Energy-efcient WDM network planning with dedicatedprotection resources in sleep mode, in: GLOBECOM 2010, 2010 IEEEGlobal Telecommunications Conference, 2010, pp. 15. doi:http://dx.doi.org/10.1109/GLOCOM.2010.5683205.

[12] S. Ricciardi, D. Careglio, F. Palmieri, U. Fiore, G. Santos-Boada, J. Sol-Pareta, Energy-oriented models for WDM networks, in: GreenNetworking 2010 Workshop (GN2010), Co-located with the 7thInternational ICST Conference on Broadband Communications,Networks, and Systems (Broadnets), 2010.

[13] M. Gupta, S. Singh, Greening of the internet, in: Proceedings of the2003 Conference on Applications, Technologies, Architectures, andProtocols for Computer Communications, SIGCOMM 03, 2003, pp.1926.

[14] J. Chabarek, J. Sommers, P. Barford, C. Estan, D. Tsiang, S. Wright,Power awareness in network design and routing, in: INFOCOM 2008,The 27th Conference on Computer Communications, IEEE, 2008, pp.457465. doi:http://dx.doi.org/10.1109/INFOCOM.2008.93.