ECO-friendly Distributed Routing Protocol for Reducing Network Energy Consumption

Transcription

1 ECO-friendly Disribued Rouing Proocol for Reducing Nework Energy Consumpion Daisuke Arai and Kiyohio Yoshihara KDDI R&D Laboraories Inc Ohara Fujimino-shi Saiama, Japan {di-arai, Absrac The growh of Inerne raffic has led o an increase in energy consumpion by nework equipmen such as rouers and swiches. Consequenly, energy consumpion is becoming a key environmenal, social, poliical and cos issue. Focusing on energy consumpion by Inerne Service Provider neworks, he energy boleneck is rouers. Thus, i is imperaive for us o reduce he energy consumpion of rouers. In his paper, we propose a new ECO-friendly disribued rouing proocol (ECO-RP) based on. In our proposal, an ECO-RP eniy periodically checks he amoun of raffic forwarded by he ECO- RP eniy. Moreover, an ECO-RP eniy floods informaion on he amoun of raffic o oher eniies such as he Link Sae Adverisemen. This allows each ECO-RP eniy o obain informaion on raffic in he nework, and o dynamically change link weighs based on he informaion. For example, when he amoun of overall nework raffic is small, he eniy changes link weighs so ha he raffic can be roued only hrough a subse of rouers, and unneeded rouers can shif ino sleep mode. Several simulaion sudies, which assume wo ECO-RP enabled neworks, are conduced. The resul shows ha he proposed proocol is able o reduce energy consumpion by abou 18.5% a maximum wihou nework congesion. I. INTRODUCTION Inerne raffic is increasing rapidly, and i has almos doubled each year since 1997 [1]. Energy consumpion by nework equipmen such as rouers and swiches o carry his raffic is also increasing [2], [3]. Consequenly, energy consumpion is becoming a key environmenal, social, poliical and cos issue. When we focus on energy consumpion by Inerne Service Provider (ISP) neworks, he energy boleneck is rouers [4], [5]. Thus, i has become imperaive o reduce he energy consumpion of rouers. Research and developmen aiming a reducing nework energy consumpion [6], [7], [8], [9], [1], [11] has been carried ou. Some research has focused on he exension of nework equipmens which change a link speed based on he uilizaion of he link [6]. Sill oher research has examined he exension of nework equipmens which can be pu ino sleep mode when no packes are buffered [7], [8], [9]. Oher researches have shown how o reconfigure a nework in order o reduce he number of rouers in use [1], [11]. To reconfigure a nework, a cenralized nework managemen scheme is shown in [11]. However, when we focus on he energy consumpion of rouers in an ISP nework, i is difficul o reduce energy consumpion solely by applying hese sysemoriened schemes [6], [7], [8], [9], because rouers in an ISP nework mus forward user raffic consanly. Furhermore, o reduce he energy consumpion of rouers by means of hese cenralized nework managemen schemes [1], [11], a nework operaor mus insall servers. In some cases, insalling servers may be difficul from a cos perspecive. Changing he nework configuraion based on he amoun of raffic has he poenial o reduce he energy consumpion of rouers, and achieving his using a disribued scheme is promising in erms of cos. In his paper, as a soluion o his, we propose a new ECO-friendly disribued rouing proocol (ECO-RP) based on Open Shores Pah Firs () [12]. We inroduce ECO-RP eniies o a nework. An ECO-RP eniy periodically checks he amoun of raffic forwarded by he eniy. In addiion, an ECO-RP eniy floods informaion on he amoun of raffic o oher eniies such as he Link Sae Adverisemen (LSA). Thereby, each ECO-RP eniy is able o obain informaion on he raffic in a nework. Each ECO- RP eniy may change link weighs dynamically based on he informaion. For example, when he amoun of overall nework raffic is small, he eniy changes link weighs so ha he raffic can be roued only hrough a subse of rouers, and unneeded rouers can shif ino sleep mode. To evaluae he performance of he proposed proocol from he viewpoin of link uilizaion, energy consumpion and maximum hop coun, we simulae nework behavior on wo neworks which are generaed by a well-known opology generaor. This paper is organized as follows: in Sec. II, we presen a nework model and is corresponding energy consumpion model, which we focus on hroughou he paper. We review relaed work in Sec. III. In Sec. IV, we propose an ECOfriendly disribued rouing proocol. We simulae he nework behavior of neworks when we deploy he proposed proocol in Sec. V. Finally, in Sec. VI, we evaluae he proposed proocol in erms of link uilizaion, energy consumpion and maximum hop coun. II. NETWORK AND ENERGY CONSUMPTION MODEL We presen a nework model and is corresponding energy consumpion model, which we focus on hroughou he paper. A. Nework model We consider an ISP nework where an rouing proocol is running. I is composed of a number of nodes (rouers). Le N denoe he se of nodes n, and L denoe he se of links /$26. c 21 IEEE 14 This paper was peer reviewed a he direcion of IEEE Communicaions Sociey by subjec maer expers for publicaion in he CNSM 21 proceedings.

2 l. We use he noion (i, j) in place of link l from node i o node j. The capaciy of link l is denoed by b l. The se of ingress-egress (IE) pairs k =(s k, k ) is denoed by K wih s k referring o he ingress node and k referring o he egress node of IE pair k. Le P k denoe he se of all possible pahs p from node s k o k. We use noion l p if link l belongs o a pah p. The raffic demand of IE pair k is denoed by d k. In addiion, he link load of link l is denoed by ll l. Each link l is associaed wih a weigh w l and he raffic is carried along shores pahs. Le Pk SP denoe he se of shores pahs (SP) from node s k o k wih respec o he link weigh w l, hus Pk SP is defined as: Pk SP = {p P k w l = min w l } (1) p P k l p l p In each node i, he incoming raffic wih he same desinaion is aggregaed and hen spli o links (i, j) ha belong o one of he shores pahs of he IE pair (i, ). Le φ ij denoe he corresponding spliing raio. Thus, φ ij refers o he fracion of overall raffic passing hrough node i and desined o node ha is forwarded on link (i, j). I is usually assumed ha hese spliing raios φ ij are equal, hus φ ij is modeled [13] as : φ 1 ij = {j :(i, j ) p for some p P(i,) SP } (2) φ ij =1 (3) j:(i,j) p for some p P SP (i,) Figure 1 shows a ypical ISP nework which is composed of a number of nodes. These nodes can be caegorized as eiher a backbone node or an access node. Backbone nodes provide an inermediae layer over access nodes. Meanwhile, access nodes connec ISP cusomers or neighboring neworks o provide he Inerne service. The raffic is generaed a he ISP cusomer or he neighboring nework and i comes ino he ISP nework. Moreover, his raffic is roued o he oher ISP cusomer or he oher neighboring nework hrough access and backbone nodes. In oher words, only he access node can be an ingress and/or egress node. The raffic demand d k changes periodically [14], [15]. Le d k () and ll l () denoe he raffic demand and he link load a ime, respecively. As an example, we show he deails of node behavior wih he raffic demand d (1,4) when all links are associaed wih he same weigh. To forward raffic, ingress node 1 calculaes shores pahs from iself o egress node 4 by using Eq. 1. Since node 1 has wo shores pahs (pah and pah 1-3-4), node 1 splis raffic demand d (1,4) o link (1, 2) and link (1, 3) according o he spliing raio φ 4 12 and φ Spliing raio φ 4 12 and φ 4 13 are calculaed by using Eq. 2 and Eq. 3. The link load ll (1,2) by he raffic demand d (1,4) is φ 4 12d (1,4) =.5d (1,4) = ll (1,2), and he link load ll (1,3) by he raffic demand d (1,4) is φ 4 13d (1,4) =.5d (1,4) = ll (1,3). In he oher node, o forward raffic, he node calculaes shores pahs from iself o he egress node, and forwards received raffic in he same way as node 1. ISP nework ISP cusomer 2 3 (2, 4) (3, 1) (3, 4) (2, 1) (1, 2) (4, 3) 1 (1, 3) (4, 2) 4 Fig. 1. neighboring nework access node (rouer) backbone node (rouer) amoun of raffic Nework model ime shores pahs (SP) from node 1 o node 4 spli raffic if i has muliple SPs B. Energy consumpion model A node is composed of one or more chassis, in which one or more line cards are insalled. Energy consumpion of a node is deermined by he sum of energy consumpion of chassis and line cards. Moreover, energy consumpion of chassis and line cards is deermined by produc ype [1]. Le cc i denoe energy consumpion of he chassis of node i, and le lcc (i,j) denoe energy consumpion of he line card of link (i, j). The nework is composed of a number of nodes, hus nework energy consumpion EC is modeled as: EC = i {cc i + j lcc (i,j) } (4) Therefore, he value of EC is deermined by he sum of energy consumpion of all chassis and line cards on a nework. Our aim is o reduce he value of EC. III. RELATED WORK Research and developmen aiming a reducing nework energy consumpion EC, has been carried ou. Sandard iniiaives include IEEE 82.3az Energy Efficien Eherne (EEE) [6]. I changes he link speed based on he uilizaion of he link by he exension of nework equipmen. Oher research has inroduced an energy-saving mode ino LAN swiches. The LAN swiches can be pu o sleep or ino an energy saving mode when no packes are buffered [2], [7], [8], [9]. However, reducing he energy consumpion of nodes solely by hese means is difficul, because every node in ISP neworks mus forward user raffic consanly, hus i is unlikely ha nodes can be pu o sleep. Some research shows ha energy consumpion could be reduced by changing he nework configuraion based on he amoun of raffic in order o reduce he number of nodes in use [1]. However, specific mechanisms for changing he nework configuraion have no been given. 21 Inernaional Conference on Nework and Service Managemen CNSM 21 15

3 A nework reconfiguraion scheme [11] has been developed. The scheme inroduces a cenral conrol server o manage he nework configuraion, and he server collecs raffic informaion from all nodes. The cenral conrol server deermines he nework configuraion in order o reduce he number of nodes in use. However, a nework operaor mus insall he cenral conrol server. In some cases, insalling servers may be difficul from a cos perspecive. For his reason, a disribued scheme wih no cenral server is needed. =2 amoun of overall NW raffic shores pahs sleep mode IV. PROPOSAL ON ECO-FRIENDLY DISTRIBUTED ROUTING PROTOCOL To reduce he EC value of Eq. 4, we focus on periodic change in raffic demand [14], [15] and propose a new ECOfriendly disribued proocol (ECO-RP). A. Proocol design ECO-RP changes he nework opology based on he amoun of overall nework raffic in order o reduce he number of acive chassis and line cards, i.e. when he amoun of overall nework raffic is small, he nework opology changes so ha raffic can be roued by a subse of nodes and heir links, and unneeded chassis and line cards can shif ino sleep mode. To change he nework opology in response o he amoun of overall nework raffic, we propose ECO-RP based on. is he mos commonly used inra-domain rouing proocol. In, raffic is roued along he shores pahs as described in Sec. II. The shores pahs are calculaed from link weigh w l which is assigned by a nework operaor. Fixed link weighs are generally used. By using fixed link weighs, he nework keeps he same opology excep when link failures occur. can dynamically re-roue around link failures; however, changing he pahs in response o he amoun of raffic is ou of he scope. Unlike, ECO-RP changes he nework opology based on he amoun of overall nework raffic. We show he behavior of an ECO-RP enabled nework in Fig. 2. ECO-RP eniies periodically check he amoun of raffic forwarded by he ECO- RP eniy. Moreover, an ECO-RP eniy floods informaion on he amoun of raffic o oher eniies. This allows each ECO- RP eniy o obain he same raffic informaion in a nework. From he raffic informaion, each ECO-RP eniy calculaes raffic rends. The raffic rend indicaes wheher he amoun of overall nework raffic is rending or. An ECO-RP eniy changes link weighs based on he raffic rend. When he raffic rend is rending ward, he ECO-RP eniy changes is link weighs so ha he raffic can be roued by a subse of rouers. Moreover, he ECO-RP eniy shifs ino sleep mode if i is no required for raffic forwarding (from he lef o righ of Fig. 2). On he oher hand, when raffic is rending wards, he ECO-RP eniy changes link weighs so ha he raffic can be roued by all rouers (from he righ o lef of Fig. 2). ime = amoun of overall NW raffic: large Fig. 2. =1 amoun of overall NW raffic : small Behavior of an ECO-RP enabled nework By he aggregaion of he raffic, link uilizaion ll l /b l is changed. If he larges link uilizaion in he nework (maximum link uilizaion) exceeds 1., nework congesion occurs, and he performance of he nework is significanly decreased. Therefore, he ECO-RP eniy should change link weighs wihou nework congesion. Moreover, in he sandard operaion of, he eniy considers he sleep node as node failure, and removes he informaion of he sleep node. To resume he sleep node, he ECO-RP eniy should disinguish beween he sleep node and node failure. B. Proocol overview We propose he following new message and daabases as addiional capabiliies. 1) Nework Sae Adverisemen (NSA). The NSA is a message. An ECO-RP eniy periodically checks he amoun of raffic forwarded by he ECO-RP eniy. Then, he ECO-RP eniy floods informaion on he amoun of raffic o oher eniies by he NSA. 2) Nework Sae Daabase (NSDB). The NSDB is a daabase which is formed by collecing NSAs from all nodes. The ECO-RP eniy calculaes he raffic rend from he NSDB. 3) Hisorical Link Sae Daabase (HLSDB). The HLSDB is a daabase which is formed by collecing Link Sae Adverisemens (LSAs). I is he same as an Link Sae Daabase (LSDB) excep ha he HLSDB conains enries of sleep nodes. The ECO-RP eniy disinguishes beween sleep nodes and node failure by he HLSDB. Each ECO-RP eniy obains he same raffic informaion in a nework by he NSA and he NSDB. From he NSDB, each ECO-RP eniy calculaes raffic rends and changes he nework opology. The acive ECO-RP eniy disinguishes beween sleep nodes and node failure by he HLSDB. Moreover, he acive ECO-RP eniy resumes a sleep node if he sleep node needs o carry he increasing raffic Inernaional Conference on Nework and Service Managemen CNSM 21

4 C. Provisioning We make he following provisioning o operae ECO-RP eniies. 1) For each ECO-RP eniy, a nework operaor provides a node class ha presens an access or backbone node. 2) For each link of he nodes, a nework operaor assigns a link weigh w l. 3) Clocks of all ECO-RP eniies in a nework are synchronized (e.g. by Nework Time Proocol (NTP)). 4) A nework operaor provides he following parameer values. All ECO-RP eniies in a nework use he same values. (Param. a) Period o check he amoun of raffic (Param. b) Period o calculae he raffic rend (Param. c) Scaling facor α o change link weighs (Param. d) Threshold β o avoid nework congesion In our proposal, a link weigh iniially assigned by a nework operaor is se o a defaul weigh w l defaul. For example, a nework operaor assigns he defaul weigh which is inversely proporional o he link capaciy [16]. When he raffic is rending ward, an ECO-RP eniy ries o avoid congesion on he nework by using he defaul weigh. In conras, when he raffic rend is rading ward, an ECO-RP eniy aggregaes raffic o a subse of rouers by changing he link weighs o reduce he number of acive chassis and line cards. D. Deails on ECO-RP We show he behavior of an ECO-RP eniy in Fig. 3. The ECO-RP eniy performs he following 5 processes: 1) Sandard operaion of : Each ECO-RP eniy has an LSA, and i includes he sae of he links, such as adjacencies and link weighs. The LSA is flooded hroughou he nework (Fig. 3 (S1)). The collecion of LSAs from all nodes forms an LSDB (Fig. 3 (S2)). An IP rouing able is calculaed from he LSDB by Eq. 1 (Fig. 3 (S3)). 2) NSA flooding and collecion: Each ECO-RP eniy moniors he amoun of raffic in each period specified by (Param. a) and floods informaion on he amoun of raffic hroughou he nework by he NSA (Fig. 3 (S4)). All ECO-RP eniies flood he NSA a he same ime. The flooded NSAs are colleced by all ECO-RP eniies. Then, an ECO-RP eniy sores hem in he NSDB (Fig. 3 (S5)). The NSDB includes he NSAs of all nodes and he hisory of he NSA changes. 3) Traffic rend calculaion: All ECO-RP eniies can ascerain he hisorical changes in he amoun of overall nework raffic from he NSDB, and hey can hen calculae he raffic rend (Fig. 3 (S6)). Afer collecing NSAs, all ECO-RP eniies calculae he rend by he slope of he linear regression (SLR) of he period specified by (Param.b). The raffic will rend eiher wards or wards. The raffic rend is calculaed by he following equaion: raffic rend = { rend, for SLR rend, for SLR < (5) ECO-RP NSDB node ime LSA adjacency link weigh w l 2 5 => => 498 Sandard operaion of Sandard operaion of NSA node 1 amoun of raffic forwarded by node 1 3 Mbps HLSDB o from LSDB o from => => -498 IP rouing able (S1) flood LSA (S2) collec LSA (S3) calculae rouing able (S4) flood NSA (S5) collec NSA (S6) calculae raffic rend (S7) change link weigh (S8) flood LSA (S9) collec LSA (S1) calculae rouing able (S11) overwrie HLSDB by LSDB (S12) swich ino sleep mode if i can Sandard operaion of (S13) remove LSDB enry of sleep node Fig. 3. (S14) change link weigh of sleep node (S15) resume sleep node if i needed ECO-RP eniies (S16) remove HLSDB enry of failure node Behavior of an ECO-RP eniy 4) Link weigh change: All ECO-RP eniies change he link weigh w l of he LSA based on he rend, he SLR, and he α (Param. c) o change he nework opology. If he raffic is rending ward, he link weigh w l is changed o he new weigh w (i,j) as below (Fig. 3 (S7)): w (i,j) = w (i,j) + SLR α, for i = access node and j mod 2 = w (i,j) SLR α, for i = access node and j mod 2 = 1 w (i,j) + SLR α, for i = backbone node and i mod 2 = w (i,j) SLR α, for i = backbone node and i mod 2 = 1 s.. 1 w (i,j) 2w (i,j) defaul 1 (6) 21 Inernaional Conference on Nework and Service Managemen CNSM 21 17

5 If he raffic is rending ward, he link weigh w l is changed o he new weigh w (i,j) as below: w (i,j) SLR α, w (i,j) = for i = access node and j mod 2 = w (i,j) SLR α, for i = backbone node and i mod 2 = s.. w (i,j) defaul w (i,j) (7) w (i,j) = w (i,j) + SLR α, for i = access node and j mod 2 = 1 w (i,j) + SLR α, for i = backbone node and i mod 2 = 1 s.. w (i,j) w (i,j) defaul (8) If he amoun of overall nework raffic becomes large when he ECO-RP eniy aggregaes raffic o a subse of nodes, nework congesion migh occur. To couner his, if he amoun of overall nework raffic is larger han β (Param. d), he link weigh w l is changed o he defaul link weigh w (i,j) defaul regardless of he raffic rend. The new weigh w (i,j) changes nework opology by he sandard operaion of (Fig. 3 (S8), (S9), (S1)). 5) Shif ino sleep mode and resume from sleep mode: Each ECO-RP eniy overwries he HLSDB wih he LSDB, when he LSDB is changed (Fig. 3 (S11)). A backbone node shifs ino sleep mode when he pahs from an access node o anoher access node are los (every IE-pair) (Fig. 3 (S12)). In his mode, he backbone node can deacivae all funcions and i can shu. In he sandard operaion of, enry of sleep node is removed from he LSDB (Fig. 3 (S13)). Only he HLSDB has he enries of sleep nodes. By comparing he HLSDB and he LSDB, each acive ECO-RP eniy can ascerain he sleep node, and change he link weigh of he sleep node on he HLSDB (Fig. 3 (S14)) by Eq. 6, Eq. 7 and Eq. 8. By calculaing shores pahs from he HLSDB, he acive ECO-RP eniy resumes he sleep node hrough he nework (e.g. by Wake-on-LAN) if he sleep node is on he pahs of every IE-pair (Fig. 3 (S15)). In addiion, if resumpion of he sleep node fails, he acive ECO-RP eniy idenifies is node as a failure node and enry of failure node is removed from he HLSDB (Fig. 3 (S16)). V. SIMULATIONS To evaluae he performance of he proposed proocol from he viewpoin of link uilizaion, energy consumpion, and maximum hop coun, we simulae nework behavior. A. Nework opologies We generae wo nework opologies, NW1 and NW2 for shor, for simulaions using he BRITE nework opology generaor [17]. The parameer seings are shown in Table I. In he generaion, we use he Waxman (random nework) model TABLE I PARAMETERS OF BRITE FOR SIMULATION opology NW1 NW2 ype Rouer only N 3 rouer model Waxman BA BRITE s α.15 - BRITE s β.2 - links/node 2 n Placemen Random access node backbone node NW1 NW2 N = 3, L =6 N = 3, L =57 he number of access nodes = N.5 =15 Fig. 4. Nework opologies for simulaion for NW1 and he Barabasi Allber (BA) model for NW2. Figure 4 shows he wo nework opologies. To caegorize each node ino eiher an access or backbone node, we calculae he Euclidean disance from he cener of he nework o each node and hen we caegorize 15 nodes (half of all nodes) as access nodes in descending order of he Euclidean disance. In he simulaions, we consider a same link capaciy b l, and a same defaul link weigh w l defaul for all links. We se link capaciy b l o 1 Gbps and defaul link weigh w l defaul o 5. B. Traffic model Only he access node can be an ingress and/or egress node as described in Sec. II. Thus he simulaed neworks can have 15 (he number of access nodes) 14 (he number of access nodes 1) = 21 IE-pairs. To consider he raffic demand of each IE-pair (sk, k), we pick wo random numbers o, r [, 1] for each node. Le o sk denoe a random number o for node sk and r k denoe a random number r for node k. Furher, for each IE-pair (sk, k) we pick a random number s (sk,k) [, 1] [18]. Each raffic demand d (sk,k) () is described as below: d (sk,k) () =γ()o sk r k s (sk,k) e δ(sk,k)/2δ (9) Here, δ(sk, k) is he Euclidean disance beween sk and k and Δ is he larges Euclidean disance beween any pair of nodes. Above, he o sk and r k model he aciviy level of he node, and a value closer o 1 suggess a node which generaes a large amoun of raffic. The disance facor e δ(sk,k)/2δ Inernaional Conference on Nework and Service Managemen CNSM 21

6 () () (lef scale) SLR (Param. b) = 3 hours (righ scale) SLR (Param. b) = 6 hours (righ scale) SLR (Param. b) = 12 hours (righ scale) raffic rend: Fig. 5. γ() for simulaion implies a greaer amoun of raffic for he closer pair of nodes. Moreover, γ() is raffic informaion o describe he amoun of raffic a ime. In he simulaions, we use raffic informaion of a real nework for γ(). Real raffic informaion of he pan-european research nework GEANT has been shown [19]. Traffic informaion for each 15-minue inerval over a 4-monh period is available. We use he raffic informaion from 12: noon Thursday 21s of April o 11:45 a.m. Saurday 23rd of April. Figure 5 shows γ() for he simulaions. In addiion, we normalize γ(), so ha he maximum link uilizaion ll l /b l is.9 when he peak raffic is forwarded by nodes o fi he real link uilizaion [19]. If he link uilizaion ll l /b l is greaer han 1., i means ha nework congesion occurs as a resul of nework opology changes. C. Parameer values for ECO-RP eniies To operae ECO-RP eniies, a nework operaor mus provide four parameer values as described in Sec. IV. We se 15 minues for (Param. a) since we use he raffic informaion obained during 15-minue inervals. In real operaion, a nework operaor would decide his parameer value o avoid frequen flooding of NSA. The ECO-RP eniy calculaes he raffic rend by he SLR for he period specified by (Param. b) as described in Sec. IV. To decide he value of (Param. b), we calculae he SLR of he raffic informaion wih he variable value of (Param. b) and we show he calculaion resul in Fig. 5. From he resul, we se his parameer o 6 hours, because frequen changes in he raffic rend can be avoided in comparison o he SLR for 3 hours, and i can deec he raffic rend more quickly han he SLR for 12 hours. In real operaion, a nework operaor can decide his parameer by conducing a preliminary survey of he raffic paerns before replacing wih ECO-RP enabled nodes. We simulae nework behavior wih variable values of (Param. c) and (Param. d), when all he nodes are replaced wih ECO-RP enabled ones. VI. EVALUATIONS The ECO-RP should change link weighs wihou nework congesion as described in Sec. IV. To evaluae his, we measure he maximum link uilizaion by he simulaion, and Fig. 6 and Fig. 7 show he resuls. Moreover, we measure he number of acive backbone nodes, acive links, and he maximum hop coun of all pahs, o evaluae he performance of ECO-RP from he viewpoin of energy consumpion and maximum hop coun, and Fig. 8 shows he resul. In his evaluaion, acive nodes and links refer o he nodes and links which forward a leas 1 bi raffic. A. Maximum link uilizaion Firs, we se α o.1 or.2 or.3 (variable) and se β o 1. (fixed). Fixed β (=1.) means ha he amoun of overall nework raffic is smaller han β over he simulaion. Figure 6 shows he maximum link uilizaion over he simulaion. Almos all he ime, he maximum link uilizaion of ECO- RP is greaer han ha of. Furhermore, he maximum link uilizaion of ECO-RP exceeds 1. when we se α o.3 in NW1 (in he final ward rend a boom-lef of Fig. 6), in a similar way, we can see nework congesion when we se α o.2 or.3 in NW2 (in he second ward rend a cener-righ and boom-righ of Fig. 6). Moreover, he increase in he maximum link uilizaion varies in response o α (he maximum link uilizaion is increased by a larger α). A large α leads o a large amoun of link weigh change as described in Sec. IV. From his, we can see ha a large amoun of link weigh change leads o greaer aggregaion of raffic. Therefore, α decides how much o aggregae he raffic. To reduce energy consumpion, a nework operaor would decide his parameer value in order o provide high aggregaion wihou nework congesion. On he oher hand, nework congesion occurs as a resul of he high aggregaion. To couner his, if he amoun of overall nework raffic is larger han β, he ECO-RP eniy changes link weigh w l o he defaul link weigh w l defaul as described in Sec. IV. To evaluae he effec of β, we se α o.3 (fixed, nework congesion occurs in boh NW1 and NW2) and se β o.6 or.7 (variable). Figure 7 shows he maximum link uilizaion over he simulaion. From he simulaion resul, we can see ha he maximum link uilizaion of ECO-RP falls below 1. in NW1 and NW2, unlike he maximum link uilizaion wih fixed β (in he final ward rend a op-lef and cenerlef of Fig. 7, in he second ward rend a op-righ and cener-righ of Fig. 7). Therefore, ECO-RP can avoid nework congesion in NW1 and NW2 by β. From he simulaion resul, he ECO-RP eniy can change link weighs wihou nework congesion in NW1 and NW2 when a nework operaor ses α o.3 and β o.6 or.7. B. Energy consumpion and maximum hop coun We evaluae ECO-RP from he viewpoin of energy consumpion and maximum hop coun when we se α o.3 and β o.6 or.7. Figure 8 shows he resuls for he simulaion. Then, we use he following measured values of Cisco GSR 21 Inernaional Conference on Nework and Service Managemen CNSM 21 19

7 NW1 maximum link uilizaion ll l / b l NW2 maximum link uilizaion ll l / b l ECO-RP=.1, =1. ECO-RP=.2, =1. ECO-RP=.3, =1. Fig. 6. Maximum link uilizaion ll l /b l (wih variable α and fixed β) NW1 maximum link uilizaion ll l / b l NW2 maximum link uilizaion ll l / b l ECO-RP=.3, =.6 ECO-RP=.3, =.7 ECO-RP=.3, =1. Fig. 7. Maximum link uilizaion ll l /b l (wih fixed α and variable β) 128 for he backbone node and Cisco 757 for he access node [1] o calculae he EC value: 1) Energy consumpion of an acive backbone node is 43 was. 2) Energy consumpion of an acive access node is 21 was. 3) Energy consumpion of an acive link for he backbone node is 7 was. 4) Energy consumpion of an acive link for he access node is 25 was. We calculae he EC value from he above measured values and he resul of Fig. 8: EC = Num. of acive backbone nodes 43 was +Num. of access nodes 21 was +Num. of acive links of backbone 7 was +Num. of acive links of access 25 was (1) Figure 9 shows he resul of EC from he simulaion resul. The average value of EC over he simulaion in NW1 is was. On he oher hand, if we use he defaul link weighs all he ime, which is equivalen o, he average value of EC is was. Consequenly, ECO-RP can reduce energy consumpion by abou 6.2% ( /1327.) on average wihou nework congesion if all he nodes are replaced wih ECO-RP enabled ones in NW1. Furhermore, he minimum value of EC of he ECO-RP enabled nework is 182. was. Therefore, ECO-RP can reduce energy consumpion by abou 18.5% (1 182./1327.) a maximum in NW1. In a similar way, we can see ha ECO-RP can reduce energy consumpion by abou 2.4% on average and 14% a maximum wihou nework congesion in NW2. In addiion, we can see ha here is a rade-off beween maximum hop coun and he degree o which EC is reduced. VII. CONCLUSIONS This paper presened a new ECO-friendly disribued rouing proocol based on o reduce he energy consumpion of rouers in ISP neworks. In he proposed proocol, he Inernaional Conference on Nework and Service Managemen CNSM 21

8 NW1 acive link (access) maximum hop coun acive link (access) acive link (backbone) acive node (backbone) maximum hop coun acive link (backbone) acive node (backbone) NW2 acive link (access) acive link (backbone) 2 acive node (backbone) 1 maximum hop coun acive link (access) acive link (backbone) 2 acive node (backbone) 1 maximum hop coun ECO-RP ECO-RP=.3, =.6 ECO-RP=.3, =.7 Fig. 8. Number of acive nodes, acive links, and maximum hop coun NW1 EC [was] ECO-RP 3 ECO-RP (=.3, (=.3, ) ) Fig. 9. average value of EC (lef scale) minimum value of EC (lef scale) average of maximum hop coun (righ scale) hop coun 7 NW2 EC [was] hop coun ECO-RP ECO-RP (=.3, (=.3, ) ) Energy consumpion and hop coun amoun of energy consumed by rouers is reduced by changing he link weigh in response o he amoun of raffic. We showed he deails of he proposed proocol. Moreover, o evaluae he performance of he proposed proocol from he viewpoin of energy consumpion, we simulaed nework behavior on wo neworks generaed by a well-known raffic generaor. The resul showed ha he proposed proocol was able o reduce energy consumpion by abou 18.5% a maximum wihou nework congesion. REFERENCES [1] A. M. Odlyzko, Inerne raffic growh: Sources and implicaions, Proceedings SPIE -Opical ransmission Sysem and Equipmen for WDM Neworking II, vol. 5247, pp.1 15, 23. [2] M. Ga and S. Singh, Greening of he Inerne, Proceedings of he ACM SIGCOMM, pp.19 26, 23. [3] K. W. Roh, F. Goldsein and J. Kleinman, Energy Consumpion by Office and Telecommunicaions Equipmen in Commercial Building Volume I: Energy Consumpion Baseline, Naional Technical Informaion Service (NTIS), U.S. Deparmen of Commerce, Springfield, VA22161, NTIS Number: PB , 22. [4] J. Baliga, K. Hinon and R. S. Tucker, Energy Consumpion of he Inerne, COIN-ACOFT, [5] J. Baliga, R. Ayre, K. Hinon and R. S. Tucker, Phoonic Swiching and he Energy Boleneck, Phoonics in Swiching, pp , 27. [6] IEEE 82.3 Energy Efficien Eherne Sudy Gro. [Online]. Available: hp://ieee82.org/3/eee sudy/index.hml [7] M. Ga, S. Grover and S. Shingh, A Feasibiliy Sudy for Power Managemen in LAN Swiches, Proceedings of 12h IEEE Inernaional Conference on Nework Proocols ICNP 24, pp , 24. [8] M. Ga and S. Shingh, Dynamic Eherne Link Shu for Energy Conservaion on Eherne Links, Proceedings of IEEE Inernaional Conference on Communicaions ICC 7, pp , 27. [9] H. Tamura, Y. Yahiro, Y. Fukuda, K. Kawahara and Y. Oie, Performance Analysis of Energy Saving Scheme wih Exra Acive Period for LAN Swiches, Proceeding of IEEE Global Telecommunicaions Conference 27 GLOBCOM 7, pp , 27. [1] J. Chabarek, J. Sommers, P. Barford, C. Esan, D. Tsiang and S. Wrigh, Power Awareness in Nework Design and Rouing, Proceedings of IEEE he 27h Conference on Compuer Communicaions INFOCOM 28, pp , 28. [11] N. Yamanaka, S. Shimizu and G. Shan, Energy Efficien Nework Design Tool for Green IP/Eherne Neworks, Proceedings of 14h Conference on Opical Nework Design and Modeling (ONDM), pp.1 5, 21. [12] RFC2328 Version 2, [13] D. D. Kouvasos ed., Traffic and Performance Engineering for Heerogeneous Neworks, River Publishers, ISBN , pp.85 17, 29. [14] A. Odlyzko, Daa Neworks are Lighly Uilized, and will Say ha Way, Review of Nework Economics, vol.2, Issue 3, 23. [15] A. Nucci, A. Sridharan and N. Taf, The Problem of Synheically Generaing IP Traffic Marices: Iniial Recommendaions, Proceedings of he ACM SIGCOMM, pp.19 32, 25. [16] Cisco. Design Guide. [Online]. Available: hp:// en/us/ech/k365/echnologies whie paper9186a894e9e.shml [17] A. Medina, A. Lakhina, I. Maa and J. Byers, BRITE: an approach o universal opology generaion, Proceedings of Modeling, Analysis and Simulaion of Compuer and Telecommunicaion Sysems, pp , 21. [18] B. Fors and M. Thor, Opimizing /IS-IS weigh in a Changing World, IEEE Jornal on Seleced Areas in Communicaions, vol.2, No.4, pp , 22. [19] S. Uhlig, B. Quoiin, J. Lepropre and S. Balon, Providing Public Inradomain Traffic Marices o he Research Communiy, ACM SIG- COMM, vol. 36, Num. 1, pp.83 86, Inernaional Conference on Nework and Service Managemen CNSM