Bandwdth Allocaton for Servce Level Agreement Aware Ethernet Passve Optcal Networks Dawd Nowak Performance Engneerng Laboratory School of Electronc Engneerng Dubln Cty Unversty Emal: nowakd@eengdcue Phlp Perry Performance Engneerng Laboratory School of Electronc Engneerng Dubln Cty Unversty Emal: perryp@eengdcue John Murphy Performance Engneerng Laboratory Department of Computer Scence Unversty College Dubln Emal: murphyj@eengdcue Abstract Passve Optcal Networks (PON) are thought to be the next step n the development of access networks and provdng broadband access n the last mle area Ethernet PONs (EPONs) are ganng the most attenton from ndustry as they offer a hghly flexble, cost effectve soluton In ths paper, we present a new approach to the problem of optmal bandwdth allocaton n EPONs We show that by movng all the access control functonalty to the Optcal Lne Termnator, a flexble soluton could be acheved whch offers full support for Servce Level Agreements and removes the burden of managng the confguraton of every Optcal Network Unt We nclude results of smulatons that show that such an approach can delver good performance n terms of average and maxmum packet delay We show that, a novel Grant Multplexng scheme can sgnfcantly reduce jtter and end-to-end delay experenced by the hgh prorty classes of traffc To prove these ponts we present results of detaled experments that were run on a C++ event drven smulator, that we haveloped 1 INTRODUCTION Wth the ncreasng popularty of the Internet, the traffc generated by domestc and small busness users has been growng constantly over the last couple of years Varous technologes have been deployed to provde broadband access to the network n the area known as the last mle As network operators strve for cost effcences, Passve Optcal Networks (PONs) seem to be the next step n the development of access networks (AN) A PON s a pont-to-multpont all optcal network wth no actve elements n the path between the sgnal source and the destnaton On the network sde there s an Optcal Lne Termnator (OLT) unt The OLT s usually placed n the local exchange and t acts as a pont of access to the Wde or Metropoltan Area Network On the customer s sde there s an Optcal Network Unt (ONU) An ONU can be placed ether n the curb, buldng or home and ts prmary task s to convert data between the optcal and electrcal domans Two protocols, Asynchronous Transfer Mode (ATM) and Ethernet, have been proposed as the transmsson protocol n PONs In recent years Ethernet based PONs (EPONs) have ganed more attenton from ndustry The archtecture of an Ethernet network s smple yet extremely effcent Interpretablty between old and new networks can easly be mantaned and legacy solutons can be used as EPON data s carred n standard Ethernet frames In a downstream transmsson, the OLT uses all the avalable bandwdth to broadcast packets through the spltter/coupler to every ONU Each ONU extracts packets from the stream based on the Logcal Lnk Identfer (LLID) that has been assgned durng the dscovery phase In the upstream drecton packets sent by an ONU can only reach the OLT as the optcal spltter prevents an ONU from recevng packets from other ONUs In order to avod collsons between frames from dfferent ONUs, at the optcal spltter, the avalable bandwdth must be shared among all ONUs The OLT s responsble for assgnng a non overlappng tme-slot to each ONU, and ONUs can only transmt durng that tme-slot Durng an off perod packets are buffered and when the tme arrves, the ONU sends a burst usng all the avalable bandwdth One of the key features of EPON networks s ther ablty to support Dfferentated Servces (DffServ) [1] archtecture and offer varous levels of qualty of servce (QoS) Generally three classes of traffc can be dstngushed: Expedted Forwardng (EF), Assured Forwardng (AF), and Best Effort (BE) EF servces (prmarly voce and vdeo) have very strct requrements and demand a constant, low end-to-end delay and jtter AF servces tend to be less senstve to packet delay but requre a guaranteed amount of bandwdth BE traffc s generated by applcatons that have no strong requrements regardng traffc propertes In ths paper we present a new approach to the problem of bandwdth assgnment We propose to keep the ONU s functonalty as smple as possble and move all necessary access control mechansms to the OLT for two man reasons: As no access control or packet schedulng s done n an ONU, varous algorthms can be deployed n the OLT wthout the need for reconfguraton of the equpment on the customer s sde It also allows for SLAs to be created, modfed and deleted durng normal network operaton An ONU wth a smple and generc archtecture s less expensve to produce and thus EPON becomes a more affordable choce
2 BACKGROUND Due to the archtecture lmtatons of EPONs, only one ONU can transmt at the tme, as sgnals from dfferent ONUs would nterfere wth each other leadng to errors at the OLT It s the OLT s responsblty to dvde the avalable bandwdth between ONUs The OLT must assgn a nonoverlappng tme-slot to every ONU and ths assgnment can be statc or dynamc In a statc allocaton a fxed amount of tme s assgned to every ONU Algorthms wth and wthout support for DffServ based on a fxed bandwdth allocaton were studed n [2], [3] Although the jtter s qute small, the avalable bandwdth s not utlzed n an effcent manner In [4], a dynamc bandwdth allocaton algorthm called Interleaved Pollng wth Adaptve Cycle Tme (IPACT) was presented Ths algorthm allocates tme-slots based on nformaton receved from ONUs durng a pollng cycle IPACT provdes statstcal multplexng and gves good bandwdth utlzaton but due to a varable pollng cycle delay senstve servces are hampered Dynamc Bandwdth Allocaton (DBA) combned wth prorty schedulng algorthms was studed n [4], [5] In both papers the OLT s responsble for grantng tme-slots to ONUs Every ONU assgns packets to dfferent queues based on ther QoS demands The man dsadvantage of ths approach s that to fully support DffServ, an ONU wll have to have knowledge about the SLA between a customer and the network provder A DBA scheme wth partal support for SLAs was studed n [6] The author proposed the statc bandwdth allocaton for EF traffc, whch could lead to neffcent bandwdth usage For non-real tme servces the bandwdth was allocated proportonally to the bandwdth requested In ths approach, a malcous user requestng a large amount of bandwdth can serously dsturb the performance of other ONUs A centralzed approach to the bandwdth allocaton was presented n [7] In ths scheme the OLT assgns bandwdth to an ONU not to a partcular queue whch s the man feature of our approach Two messages from Mult-Pont Control Protocol (MPCP) [8] sute are used n a bandwdth allocaton mechansm The REPORT message s used to report the status of queues at a partcular ONU Based on ths nformaton and the SLA, the algorthm makes a decson about the sze of a tme-slot dedcated to a partcular class of traffc n the next grantng cycle The notfcaton about the tme-slots assgnment s sent back to an ONU n the GATE message Upon recept of GRANT message every ONU allocates bandwdth to ts queues and the decson s made based only on the local knowledge and the comprehensve knowledge of utlzaton network resources s not taken nto account In ths paper we consder OLT-centrc bandwdth assgnment algorthm, where the resources are allocated to a partcular queue rather than to an ONU Ths feature makes t unque and dstngushes our approach from those already presented Octets GATE message REPORT message Octets 6 Destnaton Address Destnaton Address 6 6 Source Address Source Address 6 2 Length/Type(88 8) Length/Type(88 8) 2 2 Opcode( 2) Opcode( 2) 4 Tmestamp Tmestamp 4 1 Number of grants Number of reports 1 1 Grant map Report map 1 4 Start tme Queue Length -2 2 Length Queue Length -2 4 Start tme Queue Length -2 2 Length Queue Length -2 1 Grant map Report map 1 4 Start tme Queue Length -2 2 Length Queue Length -2 4 Start tme Queue Length -2 2 Length Queue Length -2 Fg 1 MPCP GATE and REPORT packet formats n lterature bandwdth assgnment algorthms In Fg 1 we propose the modfed verson of GRANT message that allows for per queue bandwdth assgnment 3 SLA AWARE DYNAMIC BANDWIDTH ALLOCATION ALGORITHM () The algorthm presented n ths secton s based on a proportonal allocaton of bandwdth to a class of traffc as ths provdes better utlzaton of the avalable resources The constrants agreed n the SLA are taken nto account n the second stage to ensure that classes conformng to ther SLA receve enough bandwdth The pseudocode of the scheduler s presented n Fg 2 Let Q (j) be the number of bytes reported by queue j by ONU and βk (j) s the number of bytes allocated to ths queue n step k of the algorthm Let γmn (j) and γ max(j) be the mnmum and maxmum number of bytes guaranteed to the partcular queue Let β max s the maxmum number of bytes that can be sent n a gve a gven grantng cycle such that β max = Q (j) (1),j and τ mn βmax C l τ max, (2) where τ mn and τ max s a mnmum and maxmum tme of a grantng cycle and C max s a lnk capacty In phase I, the OLT assgns number of bytes proportonally to reported queue length Q (j) β I(j) = Q (j),j Q (j) βmax (3) In the second phase the constrants gven n the SLA are appled Three dstnct stuatons have to be consdered: 1) β I (j) γ max(j) Number of assgned bytes has exceeded the maxmum The number of allocated bytes to a partcular queue s thus reduced to β II (j) = γ max(j)
Requre: Q j > for = to Onu do for j = to Queue do Req Req + Q j for = to Onu do for j = to Queue do β I(j) 1/Req Q j β max (a) PROP DBA :STEP I call PROP DBA :STEP II for = to Onu do for j = to Queue do f βi(j) γmax(j) then βii(j) γmax(j) β ex β ex + (βi(j) γmax(j)) f βi(j) γmn(j) βi(j) < γmax(j) then βii(j) βi(j) f βi(j) < γmn then f βi(j) < Q ( j) then βii(j) Q ( j) else βii(j) βi(j) :STEP III f β ex > then for = to Onu do for j = to Queue do βiii(j) βii(j)+ β ex Q (j),j Q (j) (b) OLT DBA Fg 2 Pseudocode for SLAs aware DBA scheduler 2) β I (j) γ mn (j) and β I (j) < γ max(j) Number of assgned bytes s wthn the lmts of the SLA No changes are made and β II (j) = β I (j) 3) β I (j) < γ mn (j) In a stuaton where Q (j) > β (j) number of bytes assgned s equal to β II (j) = Q (j) as t s smaller than the mnmum Otherwse no changes are made and β II (j) = β I (j) Bytes that reman unallocated durng the second step are dvded among all queues n phase three The total number of bytes allocated to a queue could be thus expressed as: β III(j) = β II(j) + β ex Q (j),j Q (j) 4 GRANT MULTIPLEXING Voce and real-tme vdeo servces are delay senstve and requre low values of jtter Snce a DBA scheme may result (4) Fg 3 EF AF BE 1 2 3 GRANTING CYCLE 1 2 3 (a) no Grant Multplexng GRANTING CYCLE (b) Grant Multplexng Bandwdth assgnment wth and wthout Grant Multplexng n consderable varaton n the assgned bandwdth n consecutve grantng cycles, t s dffcult to mantan good traffc parameters for the hgh prorty classes Here, we propose a mechansm whch could mprove the qualty of servce offered to jtter senstve classes In the GATE message,we proposed, the bandwdth granted to a partcular queue s descrbed by a two felds: start tme and length The frst nforms an ONU when t could start sendng packets from a partcular queue and the second specfes how long the tme-slot lasts for As the OLT assgns a bandwdth to a queue, not to an ONU, t s possble to grant tme-slots for hgh prorty queues from dfferent ONUs at the begnnng of the grantng cycle An example of a possble effect of Grant Multplexng mechansm on jtter reducton s shown n Fg 3 1, 2 and 3 show the varaton n tme when consecutve ONUs started sendng ther EF packets Because of the bursty nature of the AF and BE traffc, the amount of bandwdth assgned to those classes may vary consderably Ths can result n an ncrease n jtter experenced by EF classes (Fg 3(a)) Hgh prorty traffc generally demands a constant amount of bandwdth and t s less bursty n nature A sgnfcant decrease n jtter could be acheved f EF queues were assgned to consecutve tme-slots (Fg 3(b)) Improved traffc parameters comes at the prce of more bandwdth beng used for guardng ntervals that separate successve transmssons from dfferent ONUs 5 PERFORMANCE EVALUATION To measure the performance of each bandwdth allocaton algorthm we desgned an event-drven C++ based EPON smulator In our research we used 16 ONUs connected n a tree topology to a sngle OLT operatng at a speed of 1Gb/s Each ONU had three queues wth an ndependent bufferng space The guard tme between transmssons from dfferent ONUs was set to 1µs and the value of Inter-Frame Gap (IFG) between Ethernet packets was 96 bts It was shown that most network traffc (e, http, ftp and VBR servces) was best characterzed by self-smlarty and long-range dependence [9] To model a hgh prorty EF class of traffc (eg, voce applcatons) a Posson dstrbuton s
Average delay (ms) 5 4 3 2 1 wth GM Maxmum delay (ms) 4 35 3 25 2 15 1 5 wth GM (a) EF (a) EF Average delay (ms) 14 12 1 8 6 4 2 wth GM Maxmum delay (ms) 1 9 8 7 6 5 4 3 2 1 wth GM (b) AF (b) AF Fg 4 Average delay comparson Maxmum grantng cycle length 2ms Fg 5 Maxmum delay comarson Maxmum grantng cycle length 2ms generally used In our smulaton EF traffc consttuted 2% of the load The remanng load was dvded equally between AF and BE classes The length of hgh prorty traffc packets was fxed at 7 bytes The length of AF and BE packets was unformly dstrbuted between 64 and 1518 bytes The grantng cycle tme was changng dynamcally dependng on the amount of requested bandwdth We assumed that an ONU sent REPORT message for every queue at the end of the tme slot allocated to a partcular queue The grantng cycle tme was varable but t could not exceed the maxmum value We compare the performance of the algorthm aganst statc bandwdth assgnment () and proportonal dynamc bandwdth allocaton () algorthms The algorthm allocates a fxed share of bandwdth to a queue The amount allocated s based on parameters specfed upfront and the assgnment does not depend on the current length of the queue In our smulatons the bandwdth s dvded equally among all queues In, a source s assgned a bandwdth proportonally to a queue length reported by an ONU n a last REPORT message We measured the average and maxmum delay experenced by EF and AF class of traffc The BE traffc has the lowest prorty and receves the worst QoS In Fgs 4 and 5 we show the results of smulatons n a stuaton where none of the users was exceedng ts SLA The algorthm gves a very good results as long as there was enough bandwdth assgned to a partcular class of traffc (Fgs4(a) and 5(a)) but otherwse the values of average and maxmum delay are growng fast as shown n (Fgs4(b) and 5(b)) Moreover, on average gves the smallest throughput The algorthms based on proportonal bandwdth assgnment perform very well for small and medum loads For large loads above 8% the average delay s ncreasng very fast Ths s explaned by the fact that the grantng cycle tme s lmted from the top and as more packets arrve and they cannot be sent they have to be buffered untl the next cycle leadng to a dramatc ncrease n the average delay The shows much better performance than P- DBA for hgh and medum prorty traffc, whch comes at the expense of QoS receved by the best effort servces To compare the behavor of varous algorthms where some classes exceed the lmts mposed n the SLA, we ncreased transmsson rate of 8 ONUs by 1% The results of the smulatons are presented n Tables I and II The algorthm showed the best performance of all algorthms compared, although ths comes at the prce of much lower throughput as shown n Fg 6 The algorthm provded no protecton as the dfference n average delay tme for faster and slower ONUs was
TABLE I AVERAGE DELAY FOR EF CLASS FOR ONUS IN(I) AND OUTSIDE(II) OF Bandwdth utlzaton % THE LIMITS OF THE SLA TIMES IN ms (I) (II) (I) (II) (I) (II) 4 2686 19416 19876 19333 21131 19735 6 2853 19527 7835 78171 21618 2292 8 2127 19834 92451 96888 22965 2117 1 89956 96421 93852 14486 3373 31359 115 1137 19752 11183 15827 14677 13563 13 11835 24548 95653 26923 2613 1924 145 1257 15566 9618 11195 24929 23377 TABLE II AVERAGE DELAY FOR AF CLASS FOR ONUS IN(I) AND OUTSIDE(II) OF Bandwdth utlzaton % THE LIMITS OF THE SLA TIMES IN ms (I) (II) (I) (II) (I) (II) 4 3593 33114 19457 51191 36536 3382 6 35982 34472 79578 72877 37416 35348 8 36292 35431 11565 23471 38126 36923 1 14886 23787 28487 29357 4133 42131 115 19217 66639 14198 25978 15351 14986 13 1887 911 1739 3559 2736 282 145 19371 1696 1895 37377 26191 2451 small The algorthms provded good protecton and classes were always allocated the amount of bandwdth whch was agreed n the SLA Ths was especally true for samples where the total bandwdth requested was bgger than the total lnk capacty The results show that for small loads t was possble that more bandwdth was allocated to classes whch exceeded ther SLAs Ths was due to the fact that n the frst phase of the algorthm presented n ths paper, bandwdth was allocated proportonally to the reported length of queues Despte ths detal, classes that conformed to ther SLAs receved enough bandwdth In order to show the mpact of Grant Multplexng mechansm on the amount of jtter experenced by the hgh prorty classes we measured the standard devaton from the average perod of tme that elapses between the start the grantng cycle and the moment where a partcular tme-slot allocated to a hgh prorty class starts The smulaton results presented n Fg 7 show that wth Grant Multplexng mechansm the jtter could be reduced by as much as 1% Throughput 1 9 8 7 6 5 wth GM 4 4 5 6 7 8 9 1 Fg 6 Throughput comparson Jtter (ms) 1 1 1 wth Grant Multplexng 1 Fg 7 Jtter comparson 6 CONCLUSION We beleve that the ONU should be a very smple unt and all access control should be the responsblty of the OLT Ths knd of approach allows for a smooth update of the exstng polcy wthout a burden of changng the confguraton of every ONU In ths paper we proposed an algorthm based on such an approach It provdes dynamc bandwdth allocaton and access control, but all processng s done at the OLT The results of the smulatons showed that n comparson to other algorthms, the proposed algorthm offers better performance n terms of average packet delay and throughput In ths paper we also ntroduced the dea of Grant Multplexng Ths mechansm allows the OLT to assgn the bandwdth to hgh prorty queues at the begnnng of the gatng cycle The experments showed that the values of jtter experenced by hgh prorty class were consderably lower when the GM scheme was used REFERENCES [1] S Blake, D Black, M Carlson, E Daves, Z Wang, and WWess, RFC-2475 - An Archtecture for Dfferentated Servces, 1998, http://wwwetforg/rfc/rfc2475txt [2] GKramer, BMukherjee, and AMaslos, Ethernet Passsve Optcal Networks (EPON): a mssng lnk n an end-to-end optcal nternet, n Mul- Protocol Over WDM: Buldng the Next Generaton Internet, S Dxt, Ed John Wley & Sons, Inc, Mar 23, ch 8 [3] GKramer and BMukherjee, Ethernet PON: desgn and analyss of an optcal access network, Photonc Network Communcaton, vol 3, no 3, pp 37 319, July 21 [4] GKramer, BMukherjee, SDxt, YYe, and RHrth, On supportng dfferentated classes of servce n EPON-based access network, OSA Journal of Optcal Networkng, vol 1, no 8/9, pp 28 298, 22 [5] ChMAss, YYe, SDxt, and MAAl, Dynamc Bandwdth Allocaton for Qualty-of-Servce Over Ethernet PONs, IEEE Journal on Selected Areas n Communcatons, vol 21, no 9, pp 1467 1477, Nov 23 [6] Su-l Cho, Cyclc Pollng-Based Dynamc Bandwdth Allocaton for Dfferentated Classes of Servce n Ethernet Passve Optcal Networks, Photonc Network Communcatons, vol 7, no 1, pp 87 96, 24 [7] Yongqung Zhu, Maode Ma, and Tee Hang Cheng, A novel multple access scheme for Ethernet Passve Optcal Networks, n GLOBECOM 23 - IEEE Global Telecommuncatons Conference, vol 22, no 1, Dec 23, pp 2649 2653 [8] IEEE 823ah - draft, 23 [9] W Leland, M Taqqu, W Wllnger, and D Wlson, On the Self-Smlar Nature of Ethernet Traffc (extended verson), IEEE/ACM Transactons on Networkng, vol 2, no 1, pp 1 15, Feb 1994