www.ccsenet.org/mas Modern Appled Scence Vol. 4, o. 7; July 2010 A Dstrbuted Dynamc Bandwdth Allocaton Algorthm n EPO Feng Cao, Demng Lu, Mnmng Zhang, Kang Yang & Ynbo Qan School of Optoelectronc Scence and Engneerng Huazhong Unversty of Scence and Technology, Wuhan 430074, Chna atonal Engneerng Laboratory for ext Generaton Internet Access System, Wuhan 430074, Chna Wuhan atonal Laboratory for Optoelectroncs, Wuhan 430074, Chna E-mal:hellocaofeng@gmal.com Abstract EPO (Ethernet Passve Optcal etwork) s a rsng bandwdth access technology, and t could realze the comprehensve operaton access ncludng data, vdeo, and voce, wth good economc characters. IEEE 802.3ah s the ndustral standard of EPO, but t doesn t concretely regulate the uplnk bandwdth allocaton algorthm of EPO. Therefore, amng at the uplnk channel access of EPO, people have put forward varous dynamc bandwdth allocaton algorthms, but most of them belong to ntensve algorthm,.e. the dstrbuted bandwdth allocaton (DBA) algorthm runs n OLT whch s the nterceder to allocate the uplnk transmsson tme slot for OU. A new dstrbuted dynamc bandwdth allocaton algorthm (DDBA) s proposed n ths artcle, n whch OU decdes the sze of transmsson wndow based on the assstant nformaton transmtted by OLT and self queue length. The smulaton result ndcates that comparng wth IPACT (Interleaved Polng wth Adaptve Cycle Tme) (G. Kramer, 2002, P.89-107), under the hgh network load, DDBA could obvously mprove the average end-to-end tme delay and the average queue length. Keywords: EPO, DBA, Dstrbuted schedulng, Access network (A) 1. Introducton Wth the occurrence of VOIP (voce of nternet phone), vdeo conference, and contnually ncreasng data operaton, new requrements occur for the bandwdth of network. But the bottleneck of the exstng network bandwdth s n the access network, because the manstream bandwdth access mode s the technology of DSL. The DSL technology developed based on the orgnal phone network, and t doesn t ft for the transmsson of data operaton. PO s thought as one of the most hopeful techncal projects to solve ths bandwdth access bottleneck, and EPO adoptng the Ethernet protocol has larger compettve advantage, and t begns to be produced n the market. EPO could realze the comprehensve operaton access of data, voce, and vdeo, wth good economc character. Insders generally thnk that FTTH s the fnal soluton of the bandwdth access, but EPO wll be a manstream bandwdth access technology. Because of the characters of EPO network structure, the specal advantage of bandwdth access, and the natural organc combnaton wth the computer network, all experts n the whole world thnk that EPO s the optmal transmsson medum to realze three networks n one and solve the fnal one klometer of the nfo-hghway. EPO adopts the pont-to-pont network frame. The drecton from OLT to OU s the downlnk drecton, and OLT broadcasts, and OU selectvely accept the nformaton accordng to the ntenton address of frame. The drecton from OU to OLT s the uplnk drecton, and there s only one uplnk channel, and the data transmtted by dfferent OUs wll conflct. To vod that the frames sent by varous OUs conflct, a mechansm needs to harmonze the conflct nduced by varous OUs transmt data to OLT. IEEE 802.3an standard defnes mult-pont control protocol (MPCP), and t defnes fve control frames for the logn, the bandwdth request, and the bandwdth authorzaton of OU. Where, REPORT frame s the bandwdth request frame that OU transmts to OLT, and t s used to report the length of the OU buffer queue. GATE frame s the bandwdth authorzaton frame that OLT transmts to OU, and t s used to report when OU transmts data. But the standard ddn t concretely defne whch schedulng algorthm wll exceed the standard range when t s used to allocate the uplnk channel bandwdth. The uplnk bandwdth allocaton algorthm of EPO could be dvded nto the FBA and DBA. Because FBA allocate OU for fxed sze bandwdth (tme slot) each tme, so the statstcal reuse could not be used among OUs, and t could not accord wth the abruptness of the network operaton, and the bandwdth use rate s low. But DBA has not these dsadvantages, and t could not only farly allocate the uplnk bandwdth, but provde better QoS support. At present, most research works emphasze partcularly on the ntensve DBA algorthm,.e. OLT accepts the bandwdth request of OU, and compute DBA, and allocate the uplnk transmsson wndow for OU. Many artcles only mproved ths ntensve DBA algorthm, and emphasze the mprovement of the bandwdth use rate, the tme delay, and QoS. In ths artcle, a 20 ISS 1913-1844 E-ISS 1913-1852
www.ccsenet.org/mas Modern Appled Scence Vol. 4, o. 7; July 2010 new dstrbuted schedulng algorthm s proposed, and t could farly and effectvely allocate the bandwdth. In ths algorthm, the access control of the uplnk channel s ntensve, because t stll could be run n OLT, but for all OUs dstrbuted n PO n the schedulng process, OUs would decde the sze of the uplnk transmsson wndow themselves. 2. Dstrbuted DBA algorthm 2.1 Basc prncple The common character of the exstng dstrbuted DBA algorthms s that the computaton of the uplnk transmsson tme slot s accomplshed by OU, not by OLT, so OU must know the transmsson wndow nformaton of other OU. Therefore, the dstrbuted DBA algorthm proposed before modfed the frame of EPO. For example, n E. Wong s artcle (E. Wong, 2005, P.3) the network frame s smulated n LA and all OUs are conneted wth one star coupler by two fbers, and OUs could accept the frames transmtted by other OUs. SR. Sherf (SR. Sherf, 2004, P.2483-2497) changes the lght branch n EPO by a star coupler, so part of sgnal lghts come from OU could be reflected, and other OUs wll detect them. The DDBA algorthm proposed n ths artcle needs not modfy the exstng EPO frame, and only transfer the wndow nformaton transmtted by other OUs to certan one OU by the agency of OLT, and t completely accords wth the MPCP protocol regulated by IEEE 802.3ah standard, and only needs add some new felds n the REPORT frame and the GATE frame. In the algorthm of DDBA, OU decdes the sze of tle slot, and transmts the REPORT frame to OLT, and after OLT receves the REPORT frame, t transmts t to the GATE frame at one, and the authorzaton length s the tme slot requested by OU. Of course, OLT adds some extra nformaton n the GATE frame, so OU could compute the sze of the uplnk transmsson wndow. Extra nformaton ncludes a weght vector ( ),.e. the weght factors of all OUs. These weght factors are defned as the proporton of the total usng bandwdth allocated to the bandwdth of OU and one pollng cycle. OU could know the sze of other OUs transmsson wndows n last pollng cycle by the weght nformaton. When OU transmts the REPORT frame, t wll re-compute the weght factor accordng to the request bandwdth, and fll n relatve feld of the REPORT frame, and the weght nformaton from OLT wll be updated contnually. When the schedulng begns, a fxed weght wll be allocated to each OU accordng to the nformaton of SLA. The pollng mode generally ncludes the nterleaved pollng and the nterleaved pollng wth pause. In the ntensve DBA algorthm, f the nterleaved pollng mode s adopted, OLT wll mplement the DBA operaton accordng to the REPORT algorthm of each sngle one OU, and f the nterleaved pollng mode wth pause s adopted, OLT wll allocate the bandwdth after recevng the REPORT frames of all OUs. Therefore, the former mode makes aganst that OLT makes far and reasonable authorzaton, but the latter mode could ensure the justce of the bandwdth allocaton under certan condton,.e. the bandwdth use rate wll be reduced, because the tme slot wll occur n two tmes of pollng, and n ths perod, the uplnk channel could not be utlzed. For DDBA, because OU could acqure the wndow nformaton transmtted by other OUs, and OLT needs not wat the arrval of all REPORT frames, so the nterleaved pollng mode s adopted (seen n Fgure 1). 2.2 Bandwdth allocaton Here, W s the maxmum bandwdth (tme slot) whch could be allocated n one pollng cycle. One cycle s the tme expensed by OLT when t polls all OUs,.e. the sum of tme slots and the protecton tme slots among them. From the formula (1), when the sum of all OU s weght factors s 1, the bandwdth of the th OU s W. The concrete approach of the DDBA algorthm could be descrbed as follows. (1) OLT mantans one weght vector table (seen n Table 1), where denotes the weght factor of the th OU n the n th cycle. When OLT receves the REPORT frame of the th OU, t wll compare the new weght factor and the factors n the weght vector table, and f the new weght factor s bgger, t wll be used to update the value n the weght vector table. Then, OLT transmts the GATE frame to OU, and allocates the authorzaton bandwdth n next cycle, and the value of bandwdth s the sze of the REPORT frame requested by OU. If the EPO system supports QoS, the GATE wll carry the follow parameter matrx. 11( n ) ( ) 1M n = 1( n ) M Where, denotes the amount of OU, and M denotes the amount of QoS sort. In the descrpton of algorthm, Publshed by Canadan Center of Scence and Educaton 21
www.ccsenet.org/mas Modern Appled Scence Vol. 4, o. 7; July 2010 the value of M s 1. But the REPORT frame only carry the weght of sngle OU and the nformaton of QoS. (2) When OU receves the GATE frame, t wll transmt the data wth the length of R to OLT after the authorzaton tme begns. Where, R denotes the bandwdth requested by OU n the th cycle. Then OU wll compute the maxmum transmsson wndow n n+1 cycles accordng to the extra nformaton n the GATE frame, and the computaton formula s W ( n + 1) = W ( n) j = 1 j Hereby, the sze of the bandwdth requested by OU n n+1 cycles,.e. R ( n + 1) = MI( W ( n + 1), Q ), where, Q denotes the queue length of the th OU. Fnally, OU transmts the REPORT frame at the end of the authorzaton tme slot, and nforms OLT ts queue length and new weght factor. These weght factors could be obtaned by the followng formula. R ( n + 1) ( n + 1) = W j = 1 ( n) j Thus, the authorzaton to the th OU s accomplshed, and other OUs could be authorzed accordng to above process. The begnnng of the authorzaton tme s stll controlled by OLT, because only OLT know the sze of the wndow transmtted by each OU. 3. Smulaton and analyss The IPACT proposed by G. Kramer et al s an ntensve nterleaved pollng algorthm, and ts dsadvantages have been descrbed as follows. But the DDBA algorthm s a dstrbuted nterleaved pollng algorthm, t could not only ensure the bandwdth use rate theoretcally, but guarantee the justce of the bandwdth allocaton. Here, these two algorthms wll be smulated and compared. The IPACT algorthm adopts the lmtaton servce mode to confrm the sze of the transmsson wndow,.e. f the sze of the bandwdth requested by OU doesn t exceed the pre-establshed maxmum transmsson wndow, the sze of the authorzed tme slot s the bandwdth requested by OU, or else, the authorzed length s ths pre-establshed value. By adoptng the OPET network smulaton tool to model the EPO system, the network-level model s seen n Fgure 2. The model adopts the tree network topology wth one OLT and 16 OUs. The dstance from OLT to the lght branch s 5km, and the dstance from OU to the branch s n 0-15km, so RTT wll be dstrbuted n 50us to 200us evenly. Each OU node meets wth one user node whch wll produce the network operaton, and transmt t to OU. To acqure real network performance, the operaton model wth self-smlarty and correlaton s adopted (M. Crovella, 1996). The Hurst parameter wth the self-smlarty source s 0.8. Accordng to the defnton of Ethernet frame format, the lengths of the data package are n 512 bts to 12144 bts and obey the even dstrbuton. These concrete smulaton parameters nclude that the speed from the user to OU s 100Mbps, and the uplnk and downlnk channel speed s 1Gbps, and one buffer queue s set at the place of OU, and the maxmum length s 1Mbyte, and the protecton tme slot s 2ms. To smulate the performance of the smulaton algorthms n dfferent referrng loads, OU wll share the total loads evenly. For example, f the total referrng load s 1Gbps, the load referred by each OU wll be 62.5Mbps, and the operaton producton speed s 9383 packets/s. The statstcal quanttes collected by the smulaton nclude the average end-to-end tme delay, the average queue length, the uplnk bandwdth use rate, and the downlnk bandwdth use rate. The average end-to-end tme delay ncludes three aspects, (1) the pollng tme delay,.e. the tme nterval from tme when the packet arrves at OU to the tme that OU transmts the REPORT frame, (2) the authorzaton tme delay,.e. the tme nterval from the tme that OU transmts the request to the tme that OU receves the authorzaton, (3) the queue tme delay,.e. the awatng tme of the data package n the transmtter queue. The average queue length denotes the lengths of all data package n the OU buffer queue, and the unt s bt. The downlnk bandwdth use rate denotes the rato of the bandwdth consumed by the control frame and the total bandwdth. Under dfferent referrng loads, the performance data (seen n Fgure 3) of IPACT and DDBA could be obtaned after 3 seconds smulaton. The abscssa denotes the referrng loads whch are standardzed by the uplnk channel speed 1Gbps. When the abscssa s 0.7, t denotes that the whole network referrng load s 700Mbps. From the 22 ISS 1913-1844 E-ISS 1913-1852
www.ccsenet.org/mas Modern Appled Scence Vol. 4, o. 7; July 2010 Fgure 3(a), when the total referrng load s less than 0.75, the average end-to-end tme delays of two algorthms are almost same, but when the total referrng load exceeds 0.8, the average end-to-end tme delay of the DDBA algorthm s a lttle less than the tme delay of the IPACT algorthm. Especally when the load s 1, the tme delay of the IPACT algorthm s more 11ms than the tme delay of the DDBA algorthm. At the locaton of OU, the change curve of the average queue length s consstent wth the change curve of tme delay (seen n Fgure 3 (b)), and after the load exceeds 0.8, the queue lengths of two algorthms wll ncrease quckly, but the ncrease of the queue length of the DDBA algorthm s slower. From the Fgure 3 (c), when the load exceeds 0.9, the use rate of the DDS algorthm to the uplnk bandwdth wll be kept above 90%, but the IPACT algorthm only has about 88% bandwdth use rate. That could ndcate the cause that the average end-to-end tme delay and the average queue length under hgh load wll ncrease abruptly, and the use of the uplnk bandwdth wll acheve the saturaton. For the use rate of the downlnk bandwdth, two algorthms could obtan almost same result, and only when the load exceeds 0.9, the downlnk bandwdth consumed by the DDBA algorthm wll be less. 4. Conclusons Amng at the uplnk channel access of the EPO system, a new dstrbuted dynamc bandwdth allocaton algorthm s proposed n ths artcle. By the extra nformaton transmtted by OLT, varous OU could decde the sze of the uplnk transmsson wndows based on ts buffer queue length and other OU bandwdth request nformaton. The operaton process of DBA s dstrbuted n varous OUs, and OLT only asssts t. The smulaton result shows that under hgh network load, comparng wth the IPACT algorthm, ths algorthm could acqure lower end-to-end tme delay and smaller average queue length. References E. Wong, and C. Chang-Joon. (2005). Effcent dynamc bandwdth allocaton based on upstream broadcast n Ethernet passve optcal networks. Optcal Fber Communcaton Conference. o.6. P.3. G. Kramer, B. Mukherjee and G. Pesavento. (2002). Interleaved Pollng Dstrbuton Scheme n an Optcal Access etwork. Journal Photonc etwork Communcatons. o.4(1). P.89-107. M. Crovella and A. Bestavros. (1996). Self-smlarty n World Wde Web traffc: Evdence and possble causes. In Proceedngs of ACM SIGMETRICS Internatonal Conference on Measurement and Modelng of Computer Systems. Phladelpha PA, May 1996. SR. Sherf, A. Hadjantons, G. Ellnas, et al. (2004). A novel decentralzed Ethernet-based PO access archtecture for provsonng dfferentated QoS. Journal of Lghtwave Technology. o.22. P.2483-2497. Table 1. Weght vector table OU 1 OU 2 OU 3 OU ( ) ( ) ( ) 1 n 2 n 3 n 1 2 3 1 2 3 Fgure 1. Interleaved Pollng Mechansm of DDBA Publshed by Canadan Center of Scence and Educaton 23
www.ccsenet.org/mas Modern Appled Scence Vol. 4, o. 7; July 2010 Fgure 2. etwork Model of EPO Fgure 3. Performance Comparson of DDBA and IPACT 24 ISS 1913-1844 E-ISS 1913-1852