Optimizing Segment Routing using Evolutionary Computation

Transcription

1 Available online a ScienceDirec Procedia Comper Science 0 (07) 9 The h Inernaional Conference on Fre Neworks and Commnicaions (FNC-07) Opimizing Segmen Roing sing Evolionary Compaion Víor Pereira a, Migel Rocha b, Pedro Sosa a, a Cenro Algorimi / Dep. Informaics, Universiy of Minho, Porgal b Cenre of Biological Engineering, Dep. Informaics, Universiy of Minho, Porgal Absrac Segmen Roing (SR) combines he simpliciy of Link-Sae roing proocols wih he flexibiliy of Mliproocol Label Swiching (MPLS). By decomposing forwarding pahs ino segmens, idenified by labels, SR improves Traffic Engineering (TE) and enables new solions for he opimizaion of nework resorces ilizaion. This work proposes an Evolionary Compaion approach ha enables Pah Compaion Elemen (PCE) or Sofware-defined Nework (SDN) conrollers o opimize label swiching pahs for congesion avoidance while sing a he mos hree labels o configre each label swiching pah. c 06 The Ahors. Pblished by Elsevier B.V. 07 The Ahors. Pblished by Elsevier B.V. Peer-review nder responsibiliy of he Conference Program Chairs. Keywords: Segmen Roing; Traffic Engineering; Link-Sae Roing; Evolionary Compaion.. Inrodcion Traffic engineering (TE) encompasses disinc mehods o efficienly allocae nework resorces. Wih emerging new echnologies, TE mehods evolved o mee ser consrains, while increasing operaors benefis. As expeced, here are a vas nmber of proposals which aim o achieve his goal, while preserving simpliciy and scalabiliy. Since Link-Sae roing (LSR) proocols were firs proposed, nework engineers have given hem a pariclar aenion as hey inrinsically possess hose wo aribes. Indeed, hey only reqire a se of weighs assigned o each nework link o compe he forwarding pahs, shores pahs (SP), ha minimize he sm of all link weighs in he pah. Open Shores Pah Firs (OSPF) and Inermediae Sysem o Inermediae Sysem (IS-IS) are wo common LSR proocols. To achieve a beer sage of nework resorces, some OSPF/ IS-IS implemenaions se Eqal Cos Mli- Pah (ECMP). When more han one shores pah exiss o he same desinaion, ECMP enables o evenly spli raffic along he nex-hops on he pahs. B, alhogh LSR wih EMCP provides a beer disribion of raffic, i reqires a siable link weighs configraion. This configraion, sally performed by a nework adminisraor, is no easy, noably in he case of large scale neworks and heavy raffic reqiremens. A ypical sraegy consiss in seing link weighs inversely proporional o heir capaciy, InvCap, which resls in links wih greaer capaciy receiving higher volmes of raffic. This sraegy, alhogh simple, is nable o deliver an opimal raffic disribion. Forz and Thorp 4 showed ha he weigh seing problem is NP-hard, sing a convex coninos fncion o evalae neworks Corresponding ahor. Tel.: ; fax: address: pns@di.minho.p (Pedro Sosa) c 06 The Ahors. Pblished by Elsevier B.V The Ahors. Pblished by Elsevier B.V. Peer-review nder responsibiliy of he Conference Program Chairs. 0.06/j.procs

2 Víor Pereira e al. / Procedia Comper Science 0 (07) 9 congesion based on link sage levels. Employing his objecive fncion, he ahors were able o obain weighs ha achieve a beer raffic disribion. This convex fncion, which penalizes overilized links, has been sed in many TE sdies resoring o disinc opimizaion herisics, sch as Local Search 4, Evolionary Algorihms 5, Simlaed Annealing 6, or Paricle Swarm Opimizaion 7, o opimize resorce ilizaion and, also accommodaing addiional objecives sch as delay, resilience o link failre and raffic mabiliy. Even hogh LSR/ECMP wih opimized weighs enables a good disribion of raffic load, i can no deliver an opimal sage of nework s resorces, performing a few percen off 4. The even spliing of raffic across mliple shores-pah roes performed by ECMP is an obsacle o an opimal ilizaion. In some cases, even wih opimized configraions, some nework links migh no be sed a all. To address his isse, neqal load balancing of raffic among ogoing links needs o be considered. Many proposals were able o implemen neqal load balancing 8,9, b failed in preserving he simpliciy and scalabiliy expeced from a roing proocol. The Disribed Exponeniallyweighed Flow SpliTing (DEFT) 0, on he oher hand, is able o forward raffic along links on non-shores-pahs and indce neqal raffic spliing, wiho losing in simpliciy and scalabiliy. By solely relying on link weighs configraion, DEFT assigns flows o a nex-hop wih a probabiliy ha decreases exponenially wih he exra lengh of he pah relaive o he shores pah. The ahors experimenally showed ha DEFT was able o offer a beer ilizaion of resorces han he one provided by opimized OSPF or IS-IS wih ECMP. Addiionally, he Penalizing Exponenial Flow-spliTing (PEFT) emerges as an evolion of DEFT. The main difference is ha, in erms of flow spliing, DEFT is a link-based proocol, whereas PEFT is a pah-based proocol, ha is, PEFT exends DEFT by inclding a se of variables which encode pah informaion on he raffic spliing raios compaion. PEFT also brogh an imporan resl, as he ahors formally proved ha, by forwarding a porion of raffic along links on non-shores-pahs, i is possible o achieve an opimal ilizaion of resorces. While radiional TE approaches, as hose described earlier, focs on he improvemen of roing configraions, he novel concep of Sofware-Defined Neworking (SDN) opened opporniies for devising innovaive TE mechanisms. The SDN archiecre decoples he nework conrol and forwarding fncions ino wo separae planes, a conrol and a daa planes. This enables nework conrol o become direcly programmable and he nderlying infrasrcre o be absraced for applicaions and nework services. The formerly saic neworks, as hose wih a link-sae roing proocol, now become inelligen, responsive, and cenrally conrolled. The open applicaion programming inerface (API) provided by SDN implemenaions, sch as OpenFlow, simplifies nework operaions, as insrcions and rles can direcly be insalled on he exising flow ables. SDN also makes possible o rerieve informaion abo he nework sae and improve raffic marix esimaion procedres 4, and be more responsive o raffic variaions. Segmen roing (SR) 5 has been recenly proposed as a SDN echnology wih relevan simplificaions o he daa and conrol plane operaions. SR decoples edge-o-edge roing pahs ino smaller pahs called segmens. Analogosly o Mli-Proocol Label Swiching (MPLS) 6, SR ses a pah-label mechanism o specify he roe ha packes ms ake hrogh a nework. Ths, a roe is here niqely defined as a lis of segmen IDs (SIDs). This provides enhanced packe forwarding, while minimizing he need for mainaining awareness of mass volmes of nework sae. In fac, insead of pshing a flow enry o all he swiches in a pah, SR pshes a label sack, SIDs, ino he packe header when i arrives a he ingress node. The problem of deermining opimal configraions for SR has recenly been addressed, hrogh offline, online and raffic oblivios opimizaion algorihms for maximm link ilizaion 7, b also sing an ineger programming algorihm o assess he raffic engineering of packe neworks 8. In boh cases, resls showed ha SR does no reqire a high-label sack deph for SR o perform well. In his conex, his work proposes an Evolionary Compaion based TE approach for Single Adjacency Label Pah Segmen Roing (SALP-SR) ha, sing a he mos hree labels o configre edge-o-edge roing pahs, is able o opimize he ilizaion of nework resorces. Frhermore, we also explore he ilizaion of a spli compaion parameer o respond o variaions on raffic reqiremens and avoid congesion in pars of he nework.. Segmen Roing Traffic Engineering Segmen roing, proposed by he IETF, is a simple, scalable and highly flexible plaform. Based on he sorce roing paradigm, SR is a label-swiching echniqe ha allows edge roers o seer a packe hrogh he nework sing a lis of segmens, each idenifying a opological or service insrcion. While a service insrcion pinpoins a service, provided by a node, where a packe shold be delivered, a opological insrcion specifies a pah across

3 4 Víor Pereira e al. / Procedia Comper Science 0 (07) 9 which a packe shold be forwarded. Any edge-o-edge pah can be represened as a combinaion of opological segmens. Alhogh SR is very similar o MPLS, i neiher reqires a Resorce Reservaion Proocol (RSVP) or a Label Disribion Proocol (LDP), making i mch more flexible. The disribion of labels is assred by an exension o he Inerior Gaeway Proocol (IGP), IS-IS or OSPF, or cenrally managed by a Pah Compaion Elemen (PCE) or SDN conroller. The pah informaion reqired for a packe o raverse he nework is a lis of segmens, allocaed in he packe s header a he provider s edge (PE). In or proposal, only wo ypes of SIDs are reqired: Node SID: idenifies a nework node (assigned for each roer in he nework). I is a form of Prefix SID, a SID ha conains an IP address prefix calclaed by he IGP. A Node SID ses he node loopback address as prefix and is niqe domain wide. Traffic raveling hrogh a node segmen is roed by he IGP o he desinaion. Adjacency SID: An Adjacency Segmen represens a local segmen (inerface) o a specific SR node. Each roer assigns a locally significan segmen ID for each of is IGP adjacencies. An imporan advanage of SR is ha inermediae roers do no have o mainain any per-flow sae. The inermediae nodes only need o know he globally disribed segmen labels. Addiionally, SR sppors ECMP by design. Pahs idenified by node segmens are IGP shores-pahs, and hey inrinsically inclde all he ECMP pahs o he node. These wo feares provide remendos gains in nework scalabiliy. Exploring hese advanages, we propose an opimizaion mechanism for Segmen Roing o deliver opimal raffic disribion resoring o label pah configraions which conain a he mos hree segmen IDs, wih a he mos one adjacency SID. This work only ses segmen label pahs wih one, wo or hree SIDs o roe raffic from edge-o-edge, independenly of he size of he nework and volme of raffic. We claim ha i is possible o aain a near opimal ilizaion of he available resorces, sing only label pahs wih he following alernaives of configraion formas: -Segmen: The label pah is configred wih a niqe Node SID, he SID of he desinaion node, []; -Segmen: The label pah is configred wih a Node SID and an Adjacency SID. In his case here are wo possible configraions: ) [(s, ); ], he adjacency segmen sars a he sorce node s, and he Node SID idenifies he desinaion node ; ) [;(, )], he adjacency segmen ends a he desinaion node, and he Node SID idenifies he sar node of he adjacency segmen,. -Segmen: The label pah is configred wih a Node SID, an Adjacency SID, and a Node SID,[;(, v); ], where (, v) is he adjacency segmen.. Traffic Engineering wih Three-Segmens Roing.. Three-Segmens Roing Nework infrasrcres are consanly challenged by a growh on raffic volme ha needs o be fairly disribed. To achieve his goal, we propose an opimizaion procedre for Segmen Roing where solions encompass: ) a link weigh configraion for he link-sae IGP ha nderlies he SR nework, ) edge-o-edge SR label pahs configraion as well as ) he spli of raffic beween parallel label pahs o he same desinaion. A concepal represenaion of he SALP-SR opimizaion model elemens is represened in Figre. The opimizaion mechanisms here proposed are implemened in a freely available framework accessible a hp://darwin.di.minho.p/neop. As depiced in Figre, he opimizaion is performed by a norhbond framework which inegraes a SR/Link- Sae simlaor and an Evolionary Algorihm (EA) opimizaion engine. The norhbond framework ineracs wih a PCE/SDN conroller ha collecs opology and raffic relaed informaion. Dring he opimizaion process, he disribion of raffic is achieved by resoring o DEFT or PEFT raffic spliing fncions (described in Secion.). Complemenary, he SR/Link-Sae roing simlaor is sed o evalae he qaliy of he solions provided by he EA, which are encoded as a se of link weighs, from which SR label pahs and raffic splis can also be comped. The SR/Link-Sae roing simlaor models he nework as a graph G (N, A), where N is he se of nodes and A he se of arcs. In his work, we se he normalized sm of link cos Φ proposed in 4 as primary comparison meric and main opimizaion objecive. The derivaive of Φ is defined as:

4 Víor Pereira e al. / Procedia Comper Science 0 (07) 9 5 Traffic Marices Topology Topology and Traffic informaion SR Simlaor EA Opimizaion Engine PCE/SDN Conroller Opimized Solion IGP Link Weighs SR Pahs o PE PE SR Label Pahs Desinaion Prefix PE PE SR Label Pah PE R R-R PE Traffic Spli Raio 4/5 /5 PE R PE R R PE Fig.. Opimizaion Model Concepal Archiecre f or 0 a < / f or / a < / Φ 0 f or / a = a < 9/0 70 f or 9/0 a < 500 f or a < / f or a /0 where a is a opology link and a is ilizaion nder raffic reqiremens D and a specific roing configraion. Traffic demands are represened as a marix (D) of edge-o-edge reqiremens ha can be obained from informaion gahered by he SDN conroller 4. We address he NP-hard problem of weigh-seing wih Evolionary Algorihms (EA Opimizaion engine modle of Figre ). As observed, he solion aained by he norhbond opimizaion framework will be insalled in he nework by he PCE/SDN conroller. This involves he link weighs configraion for he link-sae IGP operaion, and he SR label pahs and raffic splis for he PE nework elemens operaion... DEFT/PEFT Traffic Spliing Fncions The Disribed Exponenially-weighed Flow SpliTing (DEFT) roing proocol assigns flows o a nex-hop wih a probabiliy ha decreases exponenially wih he exra lengh of he pah when compared wih he shores pah. The disance from a node o a node, when raffic is roed hogh a node v, can be expressed as dv + w,v, where dv is he shores disance from he nex-hop v o, and w,v is he weigh of he link (, v). The exra lengh of he pah from o hrogh v, when compared o he shores pah from o, is obained by Eqaion, and denoed as h,v. h,v = dv + w,v d () P ( ) Γ ( ) h h,v,v = (,i) A Γ ( ), () h,i Γ ( ) e h h,v p, if d,v = v < d 0, oherwise ( ) (4) Γ DOWN h e,v = h,v p γ, if dv < d 0, oherwise The flow proporion on he ogoing link (, v) desined o, a, is comped by Eqaion, where he exponenial fncion Γ, Eqaion 4, maps he exra lengh h,v of a pah ino [0; ]. For a node in a shores pah from s o, if all possible nex-hops are also on shores pahs, all he exra lenghs h,v are eqal o 0, and he porions of raffic P ( ) h,v are provided by ECMP. The condiion dv < d (Eqaion 4) ensres ha raffic is forwarded owards he desinaion, prevening loops. The parameer p was hogh as a consan o scale he penalizing fncion Γ o a range [w min, w max ] of possible link weighs. In his work we consider a defal p vale of and explicily define an ilizaion hreshold of % beyond () (5)

5 6 Víor Pereira e al. / Procedia Comper Science 0 (07) 9 0 % 50% 6 % 4% % 4% v v v 50% v v 4% v v 4% v 0 v 0 9 v v v p = p = 0 (a) (b) (c) (d) Fig.. DEFT Traffic Spliing Example SP from s o s SP from s o SP from v o Adjacency Segmen v 5 5 s -v 4 5 s 5 (a) (b) (c) Fig.. Three-Segmens Roing Example which a link isn sed o forward raffic. We also will consider scenarios, Sec. 4., where he p vales are disinc variables assigned o each node, ha can be opimized o improve raffic disribion. Figre provides an example for DEFT raffic spliing and also highligh he effec of disinc vales of p. In he scenario shown in Figre, raffic needs o be roed from node o node. Considering he link weighs configraion from Figre (a), he nex-hops on a shores pah from o are v and v. Therefore, sing ECMP, raffic will be eqally spli and forwarded o boh nex-hops (Figre (b)). However, DEFT enables roers o forward raffic on non-shores pahs. All adjacen nodes on non-shores pahs ha are a a shores disance less han d(, )(= ) of will be considered nex-hops, as i is he case wih v, d(v, ) = 0. For p =, 6% of raffic for, arriving a, will be forwarded o v (Figre (c)). For p = 0, his percenage increases o %, and v receives almos one hird of he raffic (Figre (d)). The remaining raffic is eqally spli beween he shores pahs nex-hops. The Penalizing Exponenial Flow-spliTing (PEFT) mechanism adds pah informaion, as real vales γ, o he DEFT spliing fncion, Eqaion 5. In his work we considered a Downward PEFT implemenaion, where he γ vales are obained by solving a linear eqaion sysem for each desinaion... Three-Segmens Roing Traffic Engineering Figre (a) presens par of a nework opology, configred wih OSPF or IS-IS, ha corresponds o he shores pah ree from s o, and he corresponding raffic splis. The raffic ha arrives a is / (/ + / /) of he raffic wih desinaion, which wold be forwarded along he link (, ). If (, ) is nable o accommodae sch volme, packe losses and increased delays wold occr. To preven sch scenario, and resoring o SR, a porion of he raffic can be forwarded along adjacency segmens ha do no belong o any shores pah from o, (a). By opimizing he disribion of raffic, sing PEFT (or DEFT) and he EA opimizaion engine (of Figre ), a (near) opimal solion for raffic disribion can be obained, and ranslaed ino an IGP link weighs configraion and raffic spliing raios, Figre (b). To accomplish he SR TE solion, wo label pahs need o be configred: ) [], where raffic is forwarded sing he insalled IGP, ) [, (, v), ], where raffic is forwarded from s o by he IGP,

6 Víor Pereira e al. / Procedia Comper Science 0 (07) 9 7 across he adjacency segmen (, v), and from v o by he IGP along he exisen shores pahs. In a sense, his can be ndersood as a load balancing of raffic beween shores pahs. F (s,,, v) = P ( h,v), = s P ( ) ( h,v p P s, (i, j) p P ( )) h (6) i, j, s F (s, ) = F (s,,, v) (7) (,v) A s, The fracion of raffic o be assigned o each label pah beween s and can be easily comped sing Eqaions 6, for he label pah conaining an adjacency segmen (, v), and 7, for a label pah ha only conains he desinaion node SID. In he eqaions, P s, represens he se of all shores pahs from s o, and A s, is he se of all non shores pah adjacency segmens on he shores pahs from s o. By applying boh eqaions o he provided example, he raffic spliing raios are 4/5 for he label pah wih he adjacency SID, and /5 o he pah wih only he node SID of, Figre (c). I is imporan o highligh ha here are differences beween he devised SR oriened proposal and he original DEFT/PEFT mehods. Conrary o or proposal, he las do no ensre ha a flow raverses a mos one non shores pah link, and limaely, a flow cold ravel solely sing non-shores pahs. The solions provided by his proposal are expeced o have eqivalen qaliy o hose offered by DEFT or PEFT when he raffic necessiies are well disribed. Frhermore, SR inrodces advanages, sch as less nework sae informaion o be mainained in inermediae roers, and he abiliy o perform per-flow TE wiho any change o he insalled configraion. 4. Experimens and Resls 4.. Experimens Seings The validaion of he mehod sed randomly generaed nework opologies (Table ), varying in size and degree. For each, we prodced raffic demand marices D wih disinc levels α of expeced link ilizaion. For each pair of nodes (s, ), he raffic needs are modeled by Eqaions 8 and 9, where R is a random nmber in [0, ], ds(s, ) he Eclidean disance beween boh nodes, c a he average capaciy of all links, A he nmber of links and hops(s, ) he minimm nmber of hops beween s and. Ths, raffic reqiremens are inversely proporional o disance. Table. Nework Topologies sed in he experimens Topology Type Nodes Links Rand0 Synheic 0 0 Rand0 4 Synheic 0 0 Rand50 Synheic Rand50 4 Synheic D (s, ) = R δ ds (s, ) hops(s, ) δ = α c a A (s,) N ds(s, ) (8) (9) The sed meric, Congesion, is he normalized sm of link cos Φ, Eqaion. I is of noice ha when his cos eqals, all loads are below / of he link capaciy, and when all arcs are exacly fll he vale of congesion is 0 /. The link weighs are inegers aken from [; 0], excep when PEFT is sed, where weighs are ranged in he [0, 00; 0, 00] inerval. All opimizaions were performed by an Evolionary Algorihm opimizaion engine. 4.. Opimizaion Resls Represenaive resls of he experimens are presened in Table. For comparison prposes, we inclde congesion vales obained wih InvCap weigh configraion, ogeher wih hose from OSPF wih ECMP opimized weighs (OSPF Op.) and for DEFT and PEFT roing proocols, also wih EA-opimized configraions. The SALP-SR resls are idenified as SR-D and SR-P according o he sed spliing fncion, respecively DEFT and PEFT. The resls show ha, alhogh he raffic disribion is no exacly he same, DEFT and SR-D solions, as well as PEFT and SR-P, have eqivalen qaliy. In all cases, SR-D and SR-P provide solions wih a beer congesion han

7 8 Víor Pereira e al. / Procedia Comper Science 0 (07) 9 OSPF/ECMP wih opimized configraions. SR-D and SR-P are able o make se of all available links which does no always happen wih OSPF/ECMP opimizaions, and conseqenly provides a beer resilience o raffic variaions. We also observe ha, in mos cases, SR-D and SR-P offer configraion solions wih eqivalen levels of congesion and he addiional compaional ime reqired by PEFT migh no be jsified. Table. Opimized Congesion Vales Topology D Level InvCap OSPF Op. DEFT SR-D PEFT SR-P Rand Rand Rand Rand Edge-o-edge raffic necessiies vary over ime, and alhogh an opimized configraion provides some level of resilience o hose variaions, i may occr ha he insalled configraions are no longer sied for neworks o accommodae he new raffic reqiremens. In some cases, by implemening per-flow TE, i is possible for a PCE o find a roing label pah wih sfficien available capaciy o ssain and seer raffic o is desinaion. Anoher possibiliy is o insall a new roing configraion which wold imply a new link weigh configraion wih conseqen impac on he nework IGP. A beer approach consiss in adjsing he raffic spliing beween parallel pahs. Alhogh his solion doesn sppor all possible variaions, i can be a good emporary fix. DEFT and PEFT se a spliing raios compaion parameer ha can be sed o improve he division of raffic beween label pahs. In Secion., we showed how differen p vales in Eqaion 4 impac he spliing of raffic. By assigning disinc p vales o each node, i is possible o aler he amon of raffic o be forwarded along adjacency segmens and shores pahs. For an insalled configraion of link weighs and SR label pahs, we propose o adjs raffic spliing beween label pahs by only opimizing he vales of p for a given new raffic marix meanwhile comped by an SDN conroller modle. Resls from his approach, ha only consider he DEFT spliing fncion, are presened in Table and inclde hree merics: congesion, he average link ilizaion (ALU) and he maximm link ilizaion (MLU). Each opology is iniially opimized for a raffic marix D 0 (Op.), and raffic variaions for wo disinc scenarios are represened as raffic marices D and D. For each of hose, we presen he measres obained wih he iniial configraion (Before) and hose obained afer he opimizaion of he p vales (Afer). For comparison, we also inclde he measres for D and D wih opimized configraions. The opimizaion was implemened sing Evolionary Algorihms where he p vales are aken wihin he range [0.0, 0.0] of real vales. By opimizing each node p vale, SALP-SR is able o address raffic demand variaions. As observed in Table, sch sraegy inrodces a clear improvemen on he raffic spliing and congesion. In a significan nmber of Table. p Vales Opimizaion Rand0 Rand0 4 Measre D 0 D D D 0 D D Op. Before Afer Op. Before Afer Op. Op. Before Afer Op Before Afer Op. Cong ALU (%) MLU (%) Rand50 Rand50 4 Measre D 0 D D D 0 D D Op. Before Afer Op. Before Afer Op. Op. Before Afer Op Before Afer Op. Cong ALU (%) MLU (%)

8 Víor Pereira e al. / Procedia Comper Science 0 (07) 9 9 scenarios he nework operaes now in an accepable working region, while before i was on a congesed siaion (vales above 0 /) or very close o i. Alhogh his sraegy always inrodces improvemens o he congesion, in some cases i s no sfficien. B, as he amon of ime reqired o obain opimized p vales is relaively shor, his approach can be evalaed in parallel wih raffic necessiies pdaes, enabling o assess if addiional measres need o be implemened. For example, he p vales for he 0 nodes opologies are opimized in less han one mine, in a reglar personal comper, and he 50 nodes opologies p vales in arond wo and a half mines. 5. Conclsions This work proposed a TE mehod, eniled as SALP-SR, able o aain near opimal resorces ilizaion on Segmen Roing neworks sing label pah configraions which se a he mos hree segmens. Using DEFT or PEFT spliing fncions o leverage raffic disribion, SR can achieve a configraion ha opimally disribes raffic on he available resorces wiho compromising simpliciy and scalabiliy. By providing an SR nework wih an opimized congesion, addiional TE asks can be implemened wiho he brden of performing complex managemen asks. We also proposed a mehod o address raffic variaions when he nework is no longer able o assre a good performance. By opimizing parameers assigned o each node, i is possible o make adjsmens o he raffic spliing beween parallel pahs and respond o a nmber of new raffic condiions, while mainaining label pah configraions. Acknowledgmens This work has been sppored by COMPETE: POCI FEDER and FCT Fndação para a Ciência e Tecnologia wihin he Projec Scope: UID/CEC/009/0. References. Moy, J.: OSPF Version. RFC 8 (Sandard). Updaed by RFC 5709 (998). Gredler, H., Goralski, W.: The Complee IS-IS Roing Proocol, Springer (005). Hopps, C.: Analysis of an Eqal-Cos Mli-Pah Algorihm. IETF RFC 99, hps://ools.ief.org/hml/rfc99 (000) 4. Forz, B.,Thorp, M.: Inerne Traffic Engineering by Opimizing OSPF Weighs. In Proceedings of IEEE INFOCOM, pages (000) 5. Pereira, V., Rocha. M.,Corez, P., Rio, M., Sosa, P.: A Framework for Robs Traffic Engineering sing Evolionary Compaion, 7h Inernaional Conference on Aonomos Infrasrcre, Managemen and Secriy (AIMS 0), Barcelona, Spain, Springer, LNCS 794, pp. - (0) 6. Ben-Amer,W., Gordin, E., Lia, B., Michel, N.: Opimizing adminisraive weighs for efficien single-pah roing. In Proc. of neworks (000) 7. Mohiddin, M., Khan, S., Engelbrech, A.: Fzzy paricle swarm opimizaion algorihms for he open shores pah firs weigh seing problem. Applied Inelligence 45,, (06) 8. Srivasava, S., Agrawal, G., Pioro, M., Medhi, D.: Deermining link weigh sysem nder varios objecives for OSPF neworks sing a Lagrangian relaxaion-based approach. IEEE Trans. New. Serv. Manage., vol., no., pp. 98, (005) 9. Sridharan, A., Gerin, R., Dio, C.: Achieving near-opimal raffic engineering solions for crren OSPF/IS-IS neworks. IEEE/ACM Trans. New., vol., no., pp. 447, (005) 0. X, D., Chiang, M., Rexford,J.: DEFT: Disribed exponenially-weighed flow spliing. Proc. IEEE Conf. Comp. Commn., pp. 779 (007).. X, D., Chiang, M., Rexford, J.: Link-sae roing wih hop-by-hop forwarding can achieve opimal raffic engineering. IEEE/ACM Trans. New. 9, 6, (0). Feamser, N., Rexford, J., Zegra, E.: The road o SDN. Qee, ():0:0 0:40 (0). McKeown, N., Anderson, T., Balakrishnan, H., Parlkar, G., Peerson, L., Rexford, J., Shenker, S., Trner, J.: OpenFlow: enabling innovaion in camps neworks. SIGCOMM Comp. Commn. Rev. 8,, (008) 4. Hark, R., Singl, D., Richerzhagen, N., Nahrsed, K., Seinmez, R.: DisTM: Collaboraive raffic marix esimaion in disribed SDN conrol planes, 06 IFIP Neworking Conference (IFIP Neworking) and Workshops, Vienna, pp (06) 5. Filsfils, C., Nainar, N. K., Pignaaro, C., Cardona, J. C., Francois, P.: The Segmen Roing Archiecre. IEEE Global Commnicaions Conference (GLOBECOM), San Diego, CA, pp. -6 (05) 6. Awdche, D., Malcolm, J., Agogba, J., O Dell, M., McMans, J.: Reqiremens for Traffic Engineering Over MPLS. RFC 70 (Informaional) (999). 7. Bhaia, R., Hao, F., Kodialam, M. S., Lakshman, T. V.: Opimized nework raffic engineering sing segmen roing. INFOCOM pp , : IEEE. ISBN: (05) 8. Moreno, E., Beghelli, A., Cgini, F.: Traffic engineering in segmen roing neworks. Comper Neworks, 4, pp. - (07)