Delay improved Media Access Control for Passive Optical Networks

Transcription

1 Copyrigh Noice c 2015 IEEE. Personal use of his maerial is permied. However, permission o reprin/republish his maerial for adverising or promoional purposes or for creaing new collecive works for resale or redisribuion o servers or liss, or o reuse any copyrighed componen of his work in oher works mus be obained from he IEEE. This maerial is presened o ensure imely disseminaion of scholarly and echnical work. Copyrigh and all righs herein are reained by auhors or by oher copyrigh holders. All persons copying his informaion are expeced o adhere o he erms and consrains invoked by each auhor s copyrigh. In mos cases, hese works may no be reposed wihou he explici permission of he copyrigh holder. Insiue of Communicaion Neworks and Compuer Engineering Universiy of Sugar Pfaffenwaldring 47, D Sugar, Germany Phone: , Fax: mail@ikr.uni-sugar.de, hp://

2 IEEE ICC 2015 SAC - Access Neworks and Sysems Delay improved Media Access Conrol for Passive Opical Neworks Sebasian Scholz Universiä Sugar, Insiue of Communicaion Neworks and Compuer Engineering, Sugar, Germany sebasian.scholz@ikr.uni-sugar.de Absrac Passive Opical Neworks (PONs) are a promising echnology for replacing oday s access neworks and o combine hem wih meropolian area neworks, because hey offer high bandwidh and cover wide disances. However, he large disances beween he endpoins inroduce he problem of a relaively large propagaion delay. Tradiional Media Access Conrol (MAC) proocols applied in oday s PONs are no able o handle hese propagaion delays efficienly. We presen a new MAC mechanism specifically ailored for operaion in Orhogonal Frequency Division Muliplexing (OFDM) based longreach PONs. Our approach combines classical polling and random access o join he benefis of boh MAC principles o reduce ransmission delays. We perform an analyical evaluaion of he combined MAC mechanism as well as simulaion sudies. The simulaion sudies show ha he new MAC approach is able o reduce he minimal ransmission delay by up o 63%. A. Moivaion I. INTRODUCTION Fuure Inerne access neworks have o provide higher bandwidh han available oday o fulfill he requiremens of upcoming applicaions. Passive Opical Neworks (PONs) are a suiable replacemen of curren access neworks, as hey can offer bandwidh up o 100 Gbi/s [1]. Besides he bandwidh, he delay of he access nework is also a resricing facor of he nework performance. Especially in he case of long-reach PONs, where he disance beween he endpoins of he access nework can be as large as 100 km, he propagaion delay alone adds up o 500 µs [2]. Applicaions like documen processing on hosed applicaions in he cloud (e. g., Google Docs or Microsof Office Web Apps) require shor ransmission delays in order o provide a good user experience. Oher applicaions like sreaming video games require high bandwidh, bu also small delays. In access neworks wih a shared medium, ransmissions from he head-end o he subscribers can be performed wihou much effor, if he nework is cenrally coordinaed by he head-end, i. e., a varian of Coordinaed Access (CA) is applied. All downlink daa have o pass he head-end where available resources of he shared medium can be assigned o ransmi he daa. Uplink ransmissions are more challenging, because he medium access of he subscribers has o be coordinaed o avoid collisions. Therefore he focus of his paper is on an appropriae Media Access Conrol (MAC) proocol for uplink ransmissions. Curren sandardized MAC proocols for PONs are designed o offer a high uilizaion of he medium, bu he delay has only a lower prioriy. For his reason we presen in his paper a new MAC proocol offering high uilizaion and shor ransmission delays. This is achieved by he combinaion of classical CA and Random Access (RA). B. Relaed Work Sandardized PON sysems, like Gigabi-PON (GPON) [3], Eherne-PON (EPON) [4] and heir successors, use polling for uplink ransmissions. This mehod requires he exchange of saus messages beween he endpoins of he PON. To gain more flexibiliy in assigning he available bandwidh o he subscribers, Dynamic Bandwidh Allocaion (DBA) algorihms are applied. In [5] he auhors compared differen DBA algorihms for EPONs wih respec o heir hroughpu and delay behavior. To reduce he delay in long-reach PONs, several improvemens for classical polling have been inroduced. Inerleaved Polling wih Adapive Cycle Time (IPACT) was one of he firs improved polling varians for EPONs [6]. In conras o classical polling, wih IPACT i is no necessary o wai unil all subscribers have repored heir resource requiremens o he head-end before resources can be allocaed. Insead, resources for one subscriber can be allocaed direcly afer he repor of his subscriber is received. Anoher concep is mulihread polling [7], which inroduces muliple concurrenly execued polling hreads o reduce he ime beween he exchange of saus messages. Anoher improved varian is real-ime polling. This varian requires an addiional uplink conrol channel o repor saus changes from he subscribers of he PON immediaely. A comparison beween polling varians showed ha boh mulihread polling and real-ime polling lead o smaller delays compared o IPACT and classical polling [2]. The scope of he EU-funded projec ACCORDANCE was o evaluae all pars of Orhogonal Frequency Division Muliplexing (OFDM) based PONs, including /15/$ IEEE 1012

3 IEEE ICC 2015 SAC - Access Neworks and Sysems 100 km f Gbi/s Power-Splier µs Signaling-Slo 150µs Daa-Slo Figure 1. Topology of he PON Figure 2. OFDM frequency ime diagram wih Codeblocks he MAC layer. Their oucome is o reuse exising MAC proocols from EPON and GPON [9]. Because he proocols for EPON and GPON are designed for a Time Division Muliplexing (TDM) operaion of he PON, hey have o be exended o suppor he addiional frequency componen of OFDM. The addiional degree of OFDM-PONs and heir advanage compared o TDM-PONs have been presened [10]. Also a MAC proocol ailored for OFDM PONs has been developed [11]. The proocol is designed as an addiion o Muli Poin Conrol Proocol (MPCP) used in EPON. Delay reducion is achieved by applying DBA and aking Qualiy of Service requiremens ino accoun. In conras o cenralized polling approaches, decenralized media access (RA) offers shorer ransmission delays a he expense of possible collisions. A MAC proocol for long-reach PONs was proposed ha is designed o reduce he delay by a RA scheme [12]. One drawback of he proposal is he requiremen of direc communicaion beween all subscribers. The idea of combining CA and RA o reduce he delay and joining he advanages of boh principles was already inroduced in [13] and [14]. Boh publicaions presen wireless neworks where raffic wih differen requiremens is ransmied. Based on he requiremens of he raffic, he appropriae MAC approach is used. RA is applied for applicaions requiring shor delays and CA is used if higher delays can be oleraed bu more daa has o be ransferred. Summarizing he relaed work i can be said ha he problem of high delays in long-reach PONs has been idenified and also soluions o reduce he delay have been proposed. Bu he proposals sill suffer from high propagaion delays. Also he benefis of decenralized MAC proocols have been idenified. However, he combinaion of CA and RA o join he benefis of boh principles is a new approach for he operaion of PONs. The remainder of he paper is organized as follows. Secion II inroduces he scenario where he presened MAC proocol is applied. Secion III gives an overview of CA varians. Secion IV presens he proposed MAC proocol. The resuls of a performance evaluaion of he proposed MAC proocol are presened in Secion V. Finally, Secion VI concludes he paper. II. SCENARIO The presened MAC proocol is designed for an OFDM based PON wih he opology shown in Fig. 1. The PON consiss of one Opical Line Terminaion () in he cenral office, up o 100 Opical Nework Unis (ONUs) a he subscriber side and a leas one opical power-splier o connec he ONUs wih he. We consider a long-reach PON which means ha he disance beween and ONUs can be up o 100 km. The resul of he long disance is a propagaion delay τ of approximaely 500 µs. We assume ha he disances beween each ONU and he are similar, so ha he difference of he propagaion delay can be negleced. One reason for his assumpion is ha he lengh of he drop-fiber (he fiber connecing an ONU wih he power splier) is much shorer han he lengh of he feeder-fiber (he fiber beween and power splier). The PON is dimensioned o carry 40 Gbi/s of raffic in boh direcions in full-duplex mode. Direc communicaion beween ONUs is no possible, which also means ha hey can no use echniques like carrier sensing o deec if hey are allowed o send. The PON is based on OFDM which offers a high flexibiliy for resource allocaion, because i is possible o allocae resources in he wo dimensions frequency and ime. The smalles possible resource ypes are Codeblocks (CBs) of fixed sizes, which are composed ou of muliple subcarriers and OFDM symbols. The ransmission of daa inside he CBs is proeced wih he help of a Forward Error Correcion (FEC) code. We assume ha i is possible o perfecly decode he daa inside a CB if no collision occurs. A collision occurs if wo or more ONUs send on he same subcarriers a he same ime. To simplify he operaion of he PON a sloed ransmission scheme of he CBs is uilized. This means ha ransmissions of CBs may only be sared a he beginning of discree imeslos. Here we assume wo differen ypes of CBs: CBs wih a size of byes. This ype is used for he normal daa ransmission and is called Daa-Codeblock (DCB). Smaller CBs wih a size of 1500byes. They are used for signaling informaion and are called Signaling-Codeblocks (SCBs). 1013

4 IEEE ICC 2015 SAC - Access Neworks and Sysems : 2 : 1 : Slo 1, 2 : Slo 3 daa leaves : 1 : 0 : Slo 1 : - : 1 : 1 : Slo 1 : Slo 2 daa leaves daa arrives a ONU daa arrives a ONU Figure 3. Principle of polling Figure 4. Principle of mulihread polling The exac sizes of he CBs are no imporan for he general idea o combine CA and RA. The presened MAC proocol only depends on ranspor mechanisms for daa and signaling informaion. The CBs are ransmied in parallel, i. e., he frequency domain is used o ransmi 10 DCBs simulaneously. We call his a daa-slo. The choice of 10 parallel DCBs is a radeoff beween good usage of available OFDM subcarriers and guard inervals in he ime domain. Wih he given bandwidh of he PON his leads o ransmission duraions of T DS = 150 µs (see Fig. 2). To suppor 100 subscribers, 100 SCBs are ransmied in parallel. This is called signaling-slo. The duraion for he ransmission is herefore T SS = 30 µs. To suppor higher numbers of subscribers more han 100 SCBs are required, which can be achieved by concaenaing several signaling-slos. As we use he frequency domain of OFDM only o ransmi several CBs in parallel, i would be also possible o apply he presened MAC proocol o oher PON echnologies like TDM-PONs. Anoher imporan aspec for he design of MAC proocols for PONs is he raffic ha has o be ransmied. Measuremens of raffic characerisics in exising access neworks show ha he carried raffic is highly bursy and flucuaing [15], [16]. We assume similar characerisics of he raffic in he PON. Therefore we design he MAC proocol o be able o handle bursy raffic efficienly. Especially a full buffer or greedy source scenario, where all subscribers of he PON wan o send as much daa as possible all he ime is no he common siuaion. III. COORDINATED ACCESS One classical operaion mode for shared media neworks is CA. A cenral saion conrols he medium access of all oher saions. Ofen his is achieved wih an implemenaion of polling, bu a oken passing mechanism is also a form of CA. The advanages of CA are he good resource uilizaion and he lack of collisions. Depending on he acual varian of CA, he waiing ime unil a saion is allowed o send can increase drasically wih increasing number of saions. A. Polling Polling is he ypical operaion mode of oday s sandardized PONs [3], [4]. In general he queries he amoun of buffered daa waiing for uplink ransmission inside he ONUs periodically. The ONUs answer he query by reporing he curren buffer level in messages. If possible, he will hen gran sufficien resources o he ONUs. Query and gran can be combined ino one message. and messages are boh ransferred in SCBs. Transmission of payload daa is performed in DATA messages in DCBs. Noe ha he polling mechanism used in he sandardized PONs differs from he polling mechanism shown in his paper. One reason is he sloed OFDM operaion in conras o TDM in EPON and GPON. Fig. 3 shows he used polling mechanism ha is suiable for a sloed sysem wih a consan polling cycle lengh. In he figure signaling-slos are colored in red and daa-slos used by ONUs 1 and 2 in yellow and blue, respecively. Unused daa-slos and daa-slos used by oher ONUs are colored in gray. The principle shown in he figure is no rue o scale o he assumed duraions of daa- and signaling-slos. Daa-slos are ransferred beween consecuive messages. The depiced and messages conain one and one for every conneced ONU. To idenify he DCBs during a polling cycle, we apply an index saring from 0 up o he number of available DCBs. This index is also used by he o assign DCBs o ONUs. The componens in he PON have o perform he following operaions in order o ransmi daa in he uplink direcion. S1 Daa arrives a he ONU and is sored in a local buffer. S2 The ONU repors he amoun of sored daa owards he in he nex signaling-slo. S3 The grans he number of needed DCBs o he ONU and sends a message. (Fig. 3 shows he assignmen of whole daa-slos for a more descripive illusraion.) S4 The ONU sends he buffered daa in he assigned DCBs. S5 The receives he daa 1014

5 IEEE ICC 2015 SAC - Access Neworks and Sysems : - : - RA: Slos 1, 2, 3, 4, 5, 6 : 0 : 1 RA: - RA: - : - : Slo 1 RA: Slos 2, 3, 4, 5, 6 daa leaves : 0 : 0 RA: 5 RA: - collision deeced : 0 : 0 RA: 4 RA: 4 : Slo 1 : Slo 2 RA: Slos 3, 4 5, 6 daa leaves daa arrives a ONU Figure 5. Principle of he combinaion of RA and CA daa arrives a ONU Figure 6. Reacion o collisions This leads o he following delay componens for uplink ransmissions. In case of overload, he overall delay increases because of necessary queuing. However, we do no consider he case of overload in he following. T 1 A variable duraion unil he ONU can send he nex message (duraion from S1 o S2). T 2 The consan duraion unil he nex message arrives from he (S2 o S3). This duraion mainly consiss of he propagaion delay. T 3 A variable duraion unil he ONU is allowed o send (S3 o S4). T 4 The consan ransmission duraion of a DCB (S4 o S5). The duraion of a polling cycle, which is defined as he ime beween he ransmissions of wo consecuive messages, should be as small as possible o offer shor delays. For opimal uilizaion of he medium he following equaion should hold: nt DS + T SS 2τ Where n denoes he number of daa-slos per polling cycle. If nt DS + T SS is smaller han 2τ, slack imes occur where he ONUs are idle and waiing for arriving messages. On he oher hand nt DS +T SS should no be much larger han 2τ, because oherwise he delay would increase. Under hese aspecs n can be calculaed as follows: 2τ TSS 1000 µs 30 µs n = = = µs T DS B. Mulihread Polling An ineresing candidae o reduce he delay especially in long-reach PONs is mulihread polling. If mulihread polling is used, hen muliple polling hreads are acive concurrenly (see Fig. 4). This reduces he waiing ime unil an ONU can repor he buffer level o he. Again he overall delay consiss of he same four pars described in he previous secion. Bu he difference is ha boh variable pars (T 1 and T 3 ) are reduced due o he fac ha here is he possibiliy o send signaling-slos more ofen. In he following we use he exreme case (n = 1), where uplink signaling-slos and daa-slos alernae. The drawback of he increased number of exchanged signaling-slos is more signaling overhead and herefore a degraded maximum uilizaion of he PON. IV. COMBINATION OF COORDINATED ACCESS AND RANDOM ACCESS This secion inroduces he main conribuion of his paper, namely he delay improved MAC mechanism. Firs we inroduce he principle, hen we presen a way how collisions can be handled efficienly. A. Principle The principle of he presened MAC procedure is o ransfer daa in unused resources in a RA fashion. Unused resources are DCBs which are lef free in he CA operaion, called RA area. Every ONU repors he buffer level as i would do in a pure CA sysem. Addiionally each ONU adds informaion abou he indices of DCBs i ried o send a DATA message in he RA area during he previous polling cycle. The operaes in he same way as i would in he CA case, bu addiionally i marks unused DCBs as RA area. Boh informaion is sen owards he ONUs in exended messages. I is possible o combine RA wih normal polling (Random-Access + Polling (RAPo)) or o combine RA wih mulihread polling (Random- Access + Mulihread Polling (RAPoMT)). During normal operaion ONUs always send newly arriving daa as soon as possible in he RA area (ransmission of in Fig. 5). If here is no DCB marked for RA, which is he case if all DCBs are used for CA, he ONU falls back o he normal polling mechanism (ransmission of in Fig. 5). To reduce he possibiliy of collisions in he RA area, ONUs are no allowed o perform RA aemps if hey already received a gran in he curren polling cycle. However, ONUs are allowed o repor daa o he and sending daa in he RA area in he same pollingcycle. B. Handling Collisions The combinaion of CA and RA leads o several challenges, caused by possible collisions beween ONUs. The deecs collisions, if i is no able o decode he received DATA messages. The resolves collisions by assigning an addiional DCB o he involved 1015

6 IEEE ICC 2015 SAC - Access Neworks and Sysems ONUs in he nex message. I knows abou he involved ONUs by comparing he index of he DCB where he collision has happened wih he informaion abou RA aemps from he messages. In he example of Fig. 6 he collision happened in slo 4 and as well as repor ha hey ried o send a DATA message in slo 4. So he knows ha a collision beween and has o be resolved. I adds an addiional DCB o boh ONUs in he nex message. Also he has o noify he ONUs abou he collision so ha hey can send he collided DATA messages again. For his noificaion we uilize an ACK mechanism, where each message conains an ACK for every correcly received DATA message. If here is no ACK for a previously send DATA message in he nex message, he ONU has o repea his DATA message. A special case of a collision happens, if wans o send daa, bu here is no RA area before he nex. Therefore i repors he curren buffer size (here assumed o correspond o one DCB) o he and sends he daa during he nex polling cycle in he RA area. In he nex cycle ransmis a DATA message in he same DCB in he RA area and a collision is deeced by he. If he would reac by assigning wo DCBs o (one for he requesed DCB from he message and one for he deeced collision) and one DCB o, i would assign one unnecessary DCB o. To solve he problem, he ONU no only repors he amoun of daa in he firs message, bu also adds he indices of he DCBs where i is going o send he DATA message in he RA area. This is possible because he informaion abou he RA area is already available from he previous message. The can use he informaion abou he planned RA aemps o avoid double grans. If he firs RA aemp of would be successful, hen would receive an unneeded gran. Compared o a pure CA scenario, his behavior is no wase of resources, because he ONU would need a gran, oo. Addiionally he unneeded gran can be used o ransmi daa ha has arrived in he meanime. If an ONU is sending muliple DATA messages in he RA area i could happen ha no all of hem arrive a he, due o collisions. Then he daa is no in order anymore and he should no forward i. Reordering can be resolved by assigning sequence numbers o every DATA messages and buffering. Bu even wih he involved buffering delay he oal delay is no larger han in he pure CA case. V. EVALUATION In his secion we compare polling, mulihread polling, RAPo and RAPoMT. As he design goal of he combined MAC approaches is a minimal delay, he evaluaion is mainly focused on he delay improvemens. The evaluaion is spli ino wo pars. In he firs par maximum and minimum values of relevan merics are derived analyically. The second par is based on a sysem level simulaion, o gain a deeper insigh in he behavior of he MAC proocols. For boh pars wo assumpions hold. Firs, and ONUs are ideal devices, which means ha hey do no inroduce any processing delays. The second assumpion is ha here is no loss in he fiber. Therefore collisons are he only cause for losses. A. Analyical Evaluaion In his subsecion we presen bounds for he achievable uilizaion and minimum and maximum limis of he delay for he evaluaed MAC mechanisms. 1) Polling: During a polling cycle here is he possibiliy o ransmi n = 7 daa-slos and one signalingslo. The maximum uilizaion can be calculaed as follows: ρ P = nt DS 1050 µs = T SS + nt DS 1080 µs = 97.2% We derive delay bounds by considering he four componens inroduced in Secion III-A. The minimum possible delay is achieved if boh variable pars (T 1 and T 3 ) vanish. Then he ransmission delay is: T P,min = T SS + nt DS + T SS + T DS + τ = 1760 µs The maximum delay omiing queuing delay is obained if T 1 and T 3 reach heir maximum possible value and can be calculaed as follows: T P,max = nt DS + T SS + nt DS + T SS + nt DS + τ = 2T SS + 3nT DS + τ = 3710 µs 2) Mulihread Polling: Mulihread polling inroduces an increased exchange of signaling-slos. Therefore he maximum achievable uilizaion is reduced o: ρ P m = T DS 150 µs = T SS + nt DS 180 µs = 83.3% The minimum possible delay is in he case of mulihread polling he same as for normal polling: T P m,min = m(t SS + nt DS ) + T SS + T DS + τ = 1760 µs m is necessary o ake he sloed operaion of muliple polling hreads ino accoun. The duraion beween an ONU sends a and receives he corresponding canno be shorer han 2τ + 2T SS. Due o he sloed srucure consising of one signaling-slo and n = 1 daa-slos, he duraion beween a and has o be an ineger muliple of T DS + nt DS. Therefore m can be calculaed as follows: 2τ + 2TSS m = = 6 T SS + nt DS 1016

7 IEEE ICC 2015 SAC - Access Neworks and Sysems Due o he muliple polling hreads, T 1 equals in he wors case he duraion of one daa-slo and one signaling-slo. The maximum delay is herefore much shorer han for he normal polling varian. T P m,max = T SS + T DS + m(t SS + nt DS ) + T SS + T DS + τ = T SS + T DS + T P m,min = 1940 µs 3) RAPo and RAPoMT: For he combined MAC mechanisms he maximum uilizaion is he same as for he underlying polling varian. In he case of RAPo wih n = 7 he uilizaion is ρ C = 97.2%. For RAPoMT (n = 1) he maximum uilizaion is ρ Cm = 83.3%. The minimum delay occurs if daa arrives shorly before he RA area and is ransmied immediaely: T C,min = T Cm,min = T DS + τ = 150 µs µs = 650 µs The maximum duraion for he ransmission of a DATA message is given by he maximum delay of he underlying polling mechanism, because in any case daa can be ransmied via CA and he access scheme guaranees ha daa in CA slos is no harmed by collisions. B. Simulaion Sudies For he sysem level simulaion a model of he PON was implemened wih he help of he even driven simulaion library IKR SimLib [17]. The model represens he PON in Fig. 1. We considered 100 ONUs, each ransmiing daa corresponding o he size of one DCB afer a negaive exponenially disribued Iner Arrival Time (IAT). The mean value of he IAT is adjused so ha he offered load of he PON can be regulaed. 100% offered load means ha he full capaciy of he PON is used for daa ransmissions, which is due o he necessary exchange of signaling informaion no achievable. In conras o he analyical consideraions, he delay in he simulaion is measured including queueing delays. Fig. 7 shows he median of he delay (sraigh line) as well as he minimal delay (doed) and he 95% quanile (dashed) of he delay over he offered load for he evaluaed mehods. Addiionally he maximum uilizaion of he MAC varians is illusraed by wo verical lines a 83.3% and 97.2% offered load. The minimal delays in he case of polling agree wih he analyical findings. The 95% quaniles of he delay say slighly below he analyical upper bounds. If he load of he PON increases o he maximum uilizaion, he delay also increases, because of shor overload siuaions. The resuls for mulihread polling also correspond o he analyical resuls. In conras o boh varians of polling, RAPo and RAPoMT offer a much lower minimum delay. Even for higher raffic loads i is sill possible o reach he minimal possible delay, whereas he delay median converges o he values of he corresponding polling counerpars. The same holds for he 95% quanile of he delay, which is always smaller han ha of he corresponding polling counerpars. The median of he delay for RAPo significanly increases for an offered load larger han 58%, whereas he delay of RAPoMT increases a higher offered loads. The reason is ha he chance for an successful ransmission direcly in he nex daa-slo in he RA area is decreased for boh varians, as more and more DCBs are used for CA ransmissions. Bu because of he shorer duraion of he polling cycle and he more frequen exchange of signaling messages in he case of RAPoMT, he overall delay is smaller han ha of RAPo. Figures 8 and 9 show he CDF of he delay for 40% and 75% offered load, respecively. As can be seen in Fig. 8 he delay for more han 80% of all ransmissions of RAPo is smaller han in he case of mulihread polling. RAPo achieves in comparison o RAPoMT for almos 70% of all ransmissions very similar delays. However, RAPoMT achieves shorer delays for all ransmissions. The CDF reveals also ha up o 65% of all ransmissions have a delay smaller han 800 µs, if RAPo is used. This means ha up o 65% of all ransmissions can be ransferred successfully in he nex available DCB (a maximum T DS waiing ime unil he nex daa-slo sars, T DS o ransmi he DCB and he propagaion delay τ). For RAPoMT even more ransmission can be done in he nex DCB. For an offered load of 75% (see Fig. 9), RAPo achieves only in almos 50% of all ransmissions a shorer delay han mulihread polling, bu says sill below he delays of polling. RAPoMT offers wih a probabiliy of 60% significan smaller delays han he mulihread polling counerpar. Bu he delay is always smaller han achieved by mulihread polling. Wih RAPo only 12% of he ransmissions can be ransferred in he nex available DCB, whereas RAPoMT can ransfer 15% of all ransmissions wihin he nex DCB. VI. CONCLUSION We have presened a new approach for MAC proocols ailored for he characerisics of long-reach PONs. We overcome he delay resricions of classical MAC proocols by combining polling wih random access. The combinaion allows o use previously unused resources o decrease he ransmission delay. The evaluaion of he mechanism showed in simulaion sudies as well as in analyical calculaions, ha he minimal delay can be reduced by 63% compared o radiional polling mechanisms. The median of he delay can be reduced by 73% in comparison o polling in ypical load siuaions. In conras o MAC proocols based compleely on RA, upper bounds for he delay can be guaraneed, because regular polling is used as a fallback in case of collisions during RA ransmissions. 1017

8 IEEE ICC 2015 SAC - Access Neworks and Sysems Delay [ms] RAPo RAPoMT Polling Mulihread Polling Median Minimum 95% Quanile Offered Load [%] Figure 7. Simulaion resuls for he ransmission delay CDF RAPo 0.2 RAPoMT Polling Mulihread Polling Delay [ms] Figure 8. CDF of he delay, 40% offered load CDF RAPo 0.2 RAPoMT Polling Mulihread Polling Delay [ms] Figure 9. CDF of he delay, 75% offered load RA collisions may lead o reordering, which can also occur wih MAC proocols based on pure CA, where reordering can happen due o losses in he fiber. Boh sources of reordering can be easily reaed by sequence numbers and buffering in he. Fuure seps include furher opimizaion of he proposed MAC proocol and a performance evaluaion wih more realisic raffic models. ACKNOWLEDGMENT This work was suppored by he German Minisry of Educaion and Research (BMBF) wihin projec ATOB under gran number 01BP1033. REFERENCES [1] Cvijeic, N. and Dayou Qian and Junqiang Hu, 100 Gb/s opical access based on opical orhogonal frequency-division muliplexing, Communicaions Magazine, IEEE, [2] Kiaei, M.S. and Fouli, K. and Scheuzow, M. and Maier, M. and Reisslein, M. and Assi, C., Delay analysis for eherne longreach passive opical neworks, in Communicaions (ICC), 2012 IEEE Inernaional Conference on, [3] ITU, Gigabi-capable Passive Opical Neworks (GPON): Physical Media Dependen (PMD) layer specificaion, Inernaional Telecommunicaions Union, [4] IEEE Sandard for Informaion echnology Local and meropolian area neworks Par 3: CSMA/CD Access Mehod and Physical Layer Specificaions Amendmen: Media Access Conrol Parameers, Physical Layers, and Managemen Parameers for Subscriber Access Neworks, IEEE Sd 802.3ah-2004, Sep [5] Michael P. McGarry and Marin Reisslein, Invesigaion of he DBA Algorihm Design Space for EPONs, J. Lighwave Technol., Jul [6] G. Kramer, B. Mukherjee, and G. Pesaveno, IPACT a dynamic proocol for an Eherne PON (EPON), Communicaions Magazine, IEEE, [7] H. Song, B.-W. Kim, and B. Mukherjee, Muli-hread polling: a dynamic bandwidh disribuion scheme in long-reach PON, Seleced Areas in Communicaions, IEEE Journal on, Feb [8] Burak Kanarci and Hussein T. Moufah, Two-sage repor generaion in long-reach EPON for enhanced delay performance, Compuer Communicaions, [9] ICT Accordance, Definiion and evaluaion of algorihms for dynamic bandwidh allocaion in ACCORDANCE, ICT Accordance, Deliverable D 4.6, June [10] Kanonakis, K. and Tomkos, I., An overview of MAC issues in OFDMA-PON neworks, in Transparen Opical Neworks (ICTON), h Inernaional Conference on, [11] Kanonakis, K. and Giacoumidis, E. and Tomkos, I., Physical- Layer-Aware MAC Schemes for Dynamic Subcarrier Assignmen in OFDMA-PON Neworks, J. Lighwave Technol., [12] Helmy, A.H. and Fahallah, H. and Abdennour, A., Decenralized media access vs. credi-based cenralized bandwidh allocaion for LR-PONs, in High Capaciy Opical Neworks and Enabling Technologies (HONET), 2011, [13] Chien-Chun Lu and Kwang-Cheng Chen, A combined polling and random access proocol for inegraed voice and daa wireless neworks, in Personal, Indoor and Mobile Radio Communicaions, Wireless Neworks - Caching he Mobile Fuure., 5h IEEE Inernaional Symposium on, [14] Buo, TheodoreV., Random Access, Reservaion and Polling Muliaccess Proocol for Wireless Daa Sysems, in Mobile Communicaions, ser. IFIP The Inernaional Federaion for Informaion Processing. Springer US, [15] G. Maier, A. Feldmann, V. Paxson, and M. Allman, On dominan characerisics of residenial broadband inerne raffic, in Proceedings of he 9h ACM SIGCOMM conference on Inerne measuremen conference, ser. IMC 09. New York, NY, USA: ACM, [16] N. Basher, A. Mahani, A. Mahani, C. Williamson, and M. Arli, A comparaive analysis of web and peer-o-peer raffic, in Proceedings of he 17h inernaional conference on World Wide Web, ser. WWW 08. New York, NY, USA: ACM, [17] IKR Simulaion and Emulaion Library, hp:// Augus