Dynamic Bandwidth Allocation Schemes in Hybrid TDM/WDM Passive Optical Networks

Transcription

1 Dynamc Bandwdth Allocaton Schemes n Hybrd TDM/WDM Passve Optcal Networks Ahmad R. Dhan, Chad M. Ass, and Abdallah Sham Concorda Insttue for Informaton Systems Engneerng Concorda Unversty, Montreal, Quebec, Canada Emal: {a dhan, ass}@cse.concorda.ca Electrcal and Computer Engneerng Department Unversty of Western Ontaro, Ontaro, Canada Emal: asham@eng.uwo.ca Abstract Ethernet Passve Optcal Networks (EPONs) are consdered the most promsng solutons for upgradng the current congested access networks to enable the delvery of broadband ntegrated servces. Although current EPON archtectures are economcally feasble, they are however bandwdth lmted. In ths paper, we dscuss a smple upgrade archtecture from EPON to WDM-PON. We present varous Dynamc Wavelength and Bandwdth Allocaton algorthms (DWBAs) that explot both nter-channel and ntra-channel statstcal multplexng n order to acheve good performance. We use extensve smulaton experments to valdate our reasonng. I. INTRODUCTION Ethernet Passve Optcal Network (EPON) represents the convergence of nexpensve and ubqutous Ethernet equpment wth low-cost fber nfrastructure. It s vewed by many as an attractve soluton for the broadband access network bottleneck and has been under ntense research actvtes recently. EPON s a pont-to-multpont access network wth no actve elements n the sgnal s path from source to destnaton [1]. It has been standardzed by the IEEE 802.3ah workng group [4] and t comprses one Optcal Lne Termnal (OLT) connected to multple Optcal Network Unts (ONUs). EPON systems deploy only one channel for downstream traffc and another channel for upstream traffc. In the downstream, Ethernet frames are broadcast by the OLT and are selectvely receved by each ONU. Alternatvely, n the upstream, multple ONUs share the same transmsson channel to transmt data and control packets to the OLT. Snce ONUs are unable to detect collson and due to the dffculty to mplement a carrer sense multple access wth collson detecton (CSMA/CD), t s necessary to desgn a mechansm that arbtrates the access of ONUs to the shared medum. Ths s acheved by desgnng a Medum Access Control (MAC) protocol to prevent collson between packets of dfferent ONUs transmttng smultaneously. Current MAC supports Tme Dvson Multplexng (TDM) [1], where each ONU s allocated a fxed or dynamc tme slot to transmt data to the OLT and each ONU buffers packets receved from dfferent subscrbers untl they are transmtted n the assgned wndow. Current EPONs effcently support up to 32 ONUs, where beyond that the performance degrades. Gven the wdespread popularty of broadband servces, the contnuous ncrease of the number of subscrbers, and the emergence of new bandwdth ntensve applcatons, the need to deploy more ONUs ( 64) becomes ndspensable. Therefore, upgradng the exstng EPON archtecture becomes necessary. There have been recently some dscussons pertanng to ths upgrade; the authors of [3] presented a new wavelength dvson multplexng (WDM) PON archtecture where multple channels are smultaneously used by the ONUs to transmt ther traffc. The authors of [6], [9] presented a new Hybrd TDM/WDM- PON archtecture named SUCCESS Standford Unversty access. In ths paper, we ntroduce a new WDM-PON archtecture and we present a possble smple mgraton from TDM-PON to TDM/WDM-PON. We present new bandwdth allocaton schemes for the hybrd WDM/TDM PON and we study ther dfferences. These new schemes enable dfferent ONUs to effcently share (both n tme and wavelength) the access network bandwdth to acheve better utlzaton. The rest of the paper s organzed as follows. Secton 2 presents the proposed WDM-PON archtecture. In secton 3, we present our bandwdth allocaton schemes and secton 4 presents analyss and comparatve study of the proposed algorthms. Fnally secton 5 concludes our work. II. WDM-PON ARCHITECTURE In our proposed archtecture, we keep the same EPON tree topology; however we ncrease the number of supported wavelengths [3]. In that context, two dfferent archtectures (A 1, A 2 ) are possble. In A 1, the set of all ONUs s dvded nto multple subsets each allocated a fxed wavelength channel for upstream transmsson. Hence, every ONU mantans a fxed transcever, whereas the OLT mantans a bank of fxed transcevers. Wthn each subset, the transmsson of dfferent ONUs s arbtrated by the OLT through ether a fxed or dynamc tme slot assgnment. Clearly, ths archtecture lmts the sharablty of dfferent wavelengths among ONUs. An alternatve A 2 (shown n Fg. 1) allows for smultaneous tme/wavelength-sharng of WDM-PON resources among all ONUs. Here, every ONU wll mantan a fast tunable laser, wth a tunng speed n the range of mcrometers and a tunng range of more than 60nm [6], [9], that enables t to tune ts upstream transmsson from one wavelength to another depng on the dynamc wavelength and bandwdth allo-

2 Fg. 1. WDM-PON Archtecture caton (DWBA) algorthm. Hence, the WDM-PON resources act as a pool and all ONUs share these resources. The sharng, however, s arbtrated by the OLT usng DWBA. Ths archtecture seems to be more favorable than the prevous, snce t ncreases the effcency of the network n terms of bandwdth utlzaton and packet delays. Nevertheless, such archtecture makes the mplementaton of the DWBA more challengng as explaned n the next secton. III. WDM/TDM DYNAMIC BANDWIDTH ALLOCATION SCHEMES Transmsson of dfferent ONUs over the shared upstream channels s typcally arbtrated by the OLT through the use of MPCP (mult-pont control protocol, IEEE 802.3ah [1]) access protocol. The OLT gathers nformaton from dfferent ONUs and dynamcally allocates bandwdth to ONUs usng REPORT and GATE messages of MPCP. The REPORT message s used by the ONU to report ts bandwdth requrements (buffer occupancy) to the OLT, whch n turn broadcasts a GATE message to that ONU, contanng the approprate transmsson wndows granted. Here, we upgrade the GATE frame to support an addtonal byte that contans the wavelength number of the channel allocated by the OLT to each ONU for transmsson. In other words, the OLT provdes each ONU wth ts approprate transmsson start tme T start, transmsson length T length and correspondng wavelength channel. MPCP does not, however, specfy any partcular bandwdth allocaton. Instead, t s desgned to facltate the mplementaton of dynamc bandwdth algorthm (DBA) [7]. One smple allocaton scheme (Statc Wavelength Dynamc Tme) SWDT reles on the smple archtecture A 1. The OLT allocates wavelengths statcally and tme slots dynamcally. The ONUs are dvded nto as many classes as the number of wavelengths, and each class of ONUs wll share a predefned wavelength. Snce the number of ONUs on each wavelength s dentfed, SWDT s run on each channel separately. Ths scheme s easy to mplement; however t underutlzes the avalable bandwdth of the offered resources snce t does not explot the nterchannel statstcal multplexng. Now unlke SWDT, the second approach reles on the second archtecture A 2 and enables the dynamc allocaton of bandwdth for dfferent ONUs n both wavelength and tme domans. Here, the OLT mantans a varable for every channel that desgnates the tme Tfree k for wavelength k where the next transmsson s possble. For every REPORT message receved from any ONU, the OLT allocates a channel wth the smallest Tfree k to ths ONU and further t also determnes the length (e.g., n bytes) of the transmsson wndow allocated to ths ONU on ths assgned channel. We refer to ths procedure as Dynamc Wavelength and Bandwdth Allocaton (DWBA) and we present three varants namely DWBA-1, DWBA-2 and DWBA-3. In these varants, the mnmum bandwdth guaranteed B MIN defned n [2], s depant from the weght assgned to each ONU based on the Servce Level Agreement (SLA) between the servce provder (SP) and users. We consder a PON network wth N ONUs. The transmsson speed of the PON s R N Mb/s. We denote T cycle as the grantng cycle whch s the tme durng whch all ONUs can transmt data or/and s REPORT messages to the OLT. We also denote T g as the guard tme that separates the transmsson wndows of two consecutve ONUs and w as the weght assgned to each ONU based on ts SLA such that N =1 w = 1. B MIN = (T cycle N T g ) R N w 8 In case of no SLA classfcaton per ONU, w = w = 1/N, and N =1 w = 1 then : B MIN = B MIN = (T cycle N T g ) R N 8 N The three varants are now ntroduced: 1)-DWBA-1 The OLT wats untl all the REPORTs are receved from all ONUs (on all channels). Upon that, the OLT runs a bandwdth allocaton [2] algorthm to determne the bandwdth and channel for every ONU. Here, f Breq B MIN where Breq s the requested bandwdth by ONU then Bassgn = B req and a GATE message s sent to ONU. Alternatvely, f Breq > B MIN then the OLT computes the excessve bandwdth resultng from the lghtly loaded ONUs and assgns to ONU a bandwdth Bassgn depng on the excess allocaton type, and ss a GATE message accordngly. There are two ways to assgn transmsson wndows va the excess bandwdth, namely Controlled Excess (CE) and Un- Controlled Excess (UE) allocaton schemes. In UE scheme, the OLT collects from the receved REPORTs all the excessve bandwdth avalable for the next cycle and assgns ths total excess unformely to all hghly loaded ONUs regardless of ther requested bandwdth. The total excess bandwdth s: B total excess = (1) (2) N B MIN Breq Breq <= B MIN (3) =1 If M s the number of overloaded ONUs, then : B excess = B total excess/m (4) The advantage of ths uncontrolled scheme s that hghly loaded ONUs are assgned enough bandwdth to satsfy ther hgh demands (assumng the excess s enough); however, f some ONUs are only slghtly hghly loaded, they are beng assgned an unfar share of the excess bandwdth that could

3 ultmately be not utlzed. Hence, the assgnment of the excess bandwdth could be controlled (.e., CE) by the OLT n order to guarantee a far bandwdth allocaton for all hghly loaded ONUs. The bandwdth assgnment under the CE scheme s as follows: B assgn = 8 >< >: Breq Breq f B req B MIN f B MIN < B req B MIN + B excess B MIN + Bexcess f B MIN < B MIN + B excess < B req (5) Let χ = {ONU } =0,...,M 1 be the set of hghly loaded ONUs. Then, Bexcess s computed as follows: { B total excess/(m ) f B MIN + (B total excess /(M )) < B req Bexcess = Breq B MIN otherwse (6a) Where the total excess Bexcess total s updated every tme Bexcess s assgned: B excess total = B excess total B excess (6b) Now for the wavelength selecton crtera, as mentoned before, the OLT mantans for every wavelength k the tme t becomes avalable for next transmsson Tfree k, k = 1,..., K where K s the total number of wavelengths n the WDM PON. The channel wth smallest Tfree k s selected for next transmsson. 2)-DWBA-2 In ths varant, upon recevng a REPORT from ONU, the OLT checks whether Breq B MIN ; n ths case, the OLT assgns on the fly a GATE to that ONU wth bandwdth Bassgn = B req. Otherwse, the OLT wats untl all the REPORTs from the other ONUs are receved and then assgns a bandwdth of Bassgn computed ether n UE or CE. The dfference here s that ONUs that are lghtly loaded can be scheduled mmedately on the partcular channel wthout watng for the rest of the ONUs to s REPORTs. Ths early allocaton wll result n mproved delay performance. 3)-DWBA-3 Ths varant s smlar to DWBA-2. Here the OLT wll always assgn on the fly a GATE to the ONU regardless of ts requested bandwdth. However the sze of the transmsson wndow s depant on the requested bandwdth. Upon recevng a REPORT from ONU, the OLT checks f Breq B MIN. If yes, then as n DWBA-2, the OLT assgns on the fly a GATE wth Bassgn = B req; otherwse, Bassgn = B MIN. Subsequently, the OLT wats untl t receves all the REPORTs from all ONUs, and collects the nformaton about the excess bandwdth from each channel as well as the number of hghly loaded ONUs (M). Agan, each hghly loaded ONU s allocated ts share of the excess bandwdth usng ether CE or UE scheme. Note, here the REPORT message s always transmtted once for each ONU n the frst assgned wndow (.e., not n the excess wndow) regardless of whether an excess bandwdth s assgned or not. That s because (1) the allocaton of the excess wndow cannot be guaranteed for a partcular ONU and (2) two REPORT messages cannot be sent because the OLT may already have done the schedulng of transmsson of other ONUs when the TABLE I DELAY MEASUREMENTS WITH UE Lnk Load SWDT DWBA-1 DWBA-2 DWBA Average Delay (seconds) frst message s receved and the second REPORT can no longer vod the frst receved one. IV. COMPARISON & ANALYSIS In ths secton we study the performance of the dfferent wavelength and bandwth allocaton algorthms presented n the prevous sectons. For that reason, we developed a WDM- PON event drven smulator n C++. The total number of ONUs N = 64, the number of wavelength channels K = 2, and the wavelength speed = 1Gbps. The guard tme s equal to 1µs, the cycle tme to 2ms and the ONU bufferng queue sze to 1MB. A subset of 32 ONUS s lghtly loaded whereas the remanng ONUs are hghly loaded. A lghtly loaded ONU generates traffc at rate of 10Mbps (load = 0.1). To model the bursty nature of Internet traffc [8], we generated self smlar traffc based on pareto dstrbuton wth a hurst H = 0.8 [5] and the sze of the generated packets vary between 64 and 1518bytes. Table 1 presents the average packet delays for SWDT, DWBA-1, DWBA-2 and DWBA- 3 all under uncontrolled excess (UE). The traffc load of a hgh loaded ONU s vared between 0.1 and 1 (.e., 10Mbps and 100M bps). SWDT clearly enables effcent bandwdth utlzaton of ndvdual channels by provdng effectve tmesharng of the bandwdth among all the ONUs of the same group. However, the algorthm does not offer any nter-channel statstcal multplexng between all the ONUs connected to the network, and therefore yelds a lower network utlzaton and hgher packet delays. Alternatvely, DWDT enables ONUs to share the network resources both n tme doman and wavelength doman. Frst, as descrbed before, DWBA-1 s straght-forward but presents some defcences. Consder two channels PON network where t1 and t2 are the tmes where each of the channels s avalable (assume t2 t1) for next transmssons. Therefore, we can compute the perod durng whch channel 1 s not beng utlzed: where: T dle 1 = (t2 t1) + K (7a) K = τ + T transmsson + T DW BA (7b) Here, τ and T transmsson are the Round Trp Tme (RTT) and the transmsson tme of a GATE message from the OLT to the ONU respectvely. T DW BA s the DWBA computaton tme.

4 (a) DWBA-2(CE) (b) DWBA-3(CE) Fg. 3. Wasted Allocated Bandwdth Packet Loss Rate Bnned Kernel Dstrbuton Estmate Packet Loss Rate vs. Lnk Load SWDT DWBA 1 DWBA 2 DWBA Lnk Load Fg. 2. Packet Loss Rate Probablty Densty Functon vs. Number of ONUs (Breq > Bmn) DWBA 2 (Un Controlled Excess) DWBA 3 (Un Controlled Excess) DWBA 2 (Controlled Excess) DWBA 3 (Controlled Excess) Number of ONUs (Breq > Bmn) Fg. 4. PDF of Number of ONUs (wth B req > B MIN ) Clearly, the upper bound of T1 dle corresponds to the maxmum of (t2 t1). Ths maxmum, n turn, corresponds to the case where the last ONU n ths cycle starts ts transmsson (on channel 2) at tme t start t1. Accordngly, the upper bound of (t2 t1) s Tassgn N (the tme wndow n seconds assgned to the last ONU). Therefore: T dle 1 T N assgn + K (8) Moreover, f t2 t1 s large (e.g., when the majorty of ONUs on channel 1 are lghtly loaded and the majorty of ONUs on channel 2 are hghly loaded), then the perod of tme where channel 1 s dle s much larger than that of (8). Overall, ths dle tme experenced by a channel results n poor bandwdth utlzaton and thus ncreased overall packet delays. DWBA- 2 and DWBA-3, on the other hand, solve ths problem of effcency by assgnng GATEs on the fly. Ths bandwdth assgnment mtgates the effects of the channel dle tme experenced by DWBA-1 and results n a better throughput and delay performance. Clearly, as shown n Table 1, at a load of 0.3(PON load,.e., on all wavelengths), the average packet delay under DWBA-2 (DWBA-3) s 17ms (14ms) better than that under DWBA-1. However, these two schemes (DWBA-2 and DWBA-3) exhbt dfferent behavors. Namely, the bandwdth allocated to a hghly loaded ONU n DWBA-2 s a complete entty (.e., mnmum guaranteed plus the excess bandwdth (controlled or uncontrolled)) whereas DWBA-3 segregates the excess bandwdth from B MIN. Ths results n two transmsson wndows beng allocated at dfferent tmes to a hghly loaded ONU n the same cycle. The mmedate mplcaton of ths transmsson segregaton s that a large packet may not ft n the frst granted wndow and gets deferred to the second granted wndow (or perhaps to a subsequent cycle), blockng other packets from beng transmtted and wastng a fracton of the allocated bandwdths. Ths mplcaton ncreases the packet delays by holdng unnecessarly these packets and ultmately resultng n ncreased buffer occupancy. On the other hand, n DWBA-2 every ONU s allocated only one transmsson wndow n the same cycle; ths wndow (combnng both B MIN and B excess ) s large enough and

5 thus mtgates the mpact of the prevous problem. Another mplcaton of allocatng two wndows n the same cycle for a hgh loaded ONU stems from the fact that the ONU reports ts buffer occupancy n the frst wndow of cycle n 1 and subsequently after some reported packets wll be transmtted n the excess wndow. The OLT now wll allocate bandwdth for cycle n based on the reported traffc from the prevous cycle. Ths wll result n grantng bandwdth more than needed snce the buffer occupancy has decreased. Let t 1,n 1 be the tme where the REPORT message s transmtted by the ONU n cycle n 1 and Q(t 1,n 1 ) be the buffer occupancy (n bytes) at tme t 1,n 1. Smlarly, let t 2,n 1 be the tme where the same ONU fnshes sng traffc n the excess wndow of cycle n 1 and Q(t 2,n 1 ) be the buffer occupancy (n bytes) at tme t 2,n 1. Here, the requested bandwdth for cycle n s Breq n = Q(t 1,n 1 ); however, an amount of Q(t 1,n 1 ) Q(t 2,n 1 ) has been transmtted already durng the excess wndow of cycle n 1. Hence, the OLT wll allocate more bandwdth than the ONU requres at tme t 2,n 1. Ths wll result n ncreased cycle tme, neffcent use of the allocated bandwdth, and fnally ncreased overall packet delays as shown n Table 1. Fgure 2 shows a comparson of the overall packet drop rate for all the allocaton algorthms. As expected, DWBA-2 exhbts a better drop rate whereas SWDT shows the hghest. Fnally, we focus our attenton on DWBA-2 and DWBA-3 to further study ther dssmlartes. Clearly, f the allocated bandwdth s not fully utlzed (as n DWBA-3), the buffer occupancy of the hgh loaded ONU wll ncrease substantally and therefore more bandwdth wll be requested for the subsequent cycle(s). Snce n our expermental setup about half of the ONUs are hghly loaded, more hgh bandwdth requests wll arrve at the OLT and each ONU wll be rewarded excess from whatever s avalable. Ths neffcency of utlzng the allocated bandwdth may occur more often, and thus may accumulate throughout the duraton of a burst and after. Alternatvely, under DWBA- 2, the behavor s conversed. Namely, the allocated bandwdth s used more effcently, and hence fewer ONUs wll be requestng addtonal bandwdth. To valdate our reasonng, we measure the probablty densty functon (pdf) of the number of ONUs wth B req B mn for both algorthms and under the two allocaton schemes of excess bandwdth (UE and CE). Clearly, as fgure 4 shows, more ONUs wll be requestng bandwdth more than B mn under DWBA-3 (both under UE and CE). However, under DWBA-2, always fewer ONUs are requestng more than B mn. Ths clearly ndcates that (1) the neffcent use of allocated bandwdth and (2) the msguded allocaton of bandwdth n DWBA-3 results n more queung of ONU s traffc and hence more ONUs requestng bandwdth larger than the mnmum guaranteed from the OLT. We also measure the number of bytes wasted n each cycle for a hghly loaded ONU under controlled excess scheme: B wasted = B alloc B sent, where B alloc and B sent are the amount of bandwdth allocated to the ONU and effectvely used by the ONU respectvely. Fgure 3 shows the results of ths experment where we plot B wasted collected n each cycle throughout the smulaton at a partcular hgh loaded ONU. Clearly, the fgure shows that, under DWBA-2, there exsts no cycle where B wasted > 1518bytes (maxmum sze of one packet). By concluson, DWBA-2 wll defer at most one packet from one transmsson cycle to the followng one; however, there s an excessve waste of allocated bandwdth under DWBA-3. Ths s largely due to the over-allocaton of unnecessary bandwdth by the OLT. V. CONCLUSION We have proposed a new WDM-PON archtecture to solve the bandwdth lmtatons of current EPON networks. A new class of allocaton algorthms (DWBAs) s proposed wheren nter- and ntra-channel statstcal multplexng are exploted to mprove the overall performance. We have shown some crtcal dssmlartes between the dfferent varants of DWBA and evaluated ther performances through extensve event drven smulatons. ACKNOWLEDGEMENTS The authors would lke to thank the anonymous revewers for ther helpful comments to ths work. REFERENCES [1] G. Kramer et al., Ethernet PON (epon): Desgn and Analyss of an Optcal Access Network, Photonc Network Communcatons, vol. 3 (2001) pp [2] C. M. Ass et al., Dynamc bandwdth allocaton for Qualty-of-Servce over Ethernet PONs, IEEE JSAC, vol. 21 (2003) pp [3] Y. Hsueh, et al., A Hghly Flexble and Effcent Passve Optcal Network Employng Dynamc Wavelength Allocaton, IEEE JLT, Vol. 23, Issue 1, 277- Jan [4] IEEE 802.3ah, Ethernet n the Frst Mle Task Force, [5] Glen Kramer, Synthetc traffc generaton, kramer/research.html. [6] F. An, et al., SUCCESS: A next-generaton hybrd WDM/TDM optcal access network archtecture IEEE/OSA Journal of Lghtwave Technology, vol. 22, no. 11, pp , Nov [7] G. Kramer et al., IPACT: A Dynamc Protocol for an Ethernet PON (EPON), IEEE Communcatons Magazne, Vol. 40, Issue 2, Feb [8] W. Leland, et al., On the Self-smlar Nature of Ethernet Traffc (Exted Verson), IEEE/ACM Transactons on Networkng, 2(1), pp Feb [9] K. S. Km, et al., Batch schedulng algorthm for SUCCESS WDM- PON Proc. of GLOBECOM 2004, Dallas, TX, USA, Nov