Efficient burst assembly algorithm with traffic prediction

Transcription

1 Efficien burs assembly algorihm wih raffic predicion Mmoloki Mangwala, Boyce B. Sigweni and Bakhe M. Nleya Deparmen of compuer science Norh Wes Universiy, Privae Bag X2046, Mmabaho, 2735 Tel: , Fax: { , ,bakhe.nleya}@nwu.ac.za Absrac In his paper we consider burs assembly algorihms and raffic predicion. The ulimae goal is o propose a new burs assembly algorihm which is based on ime-burs lengh (hybrid) hreshold wih raffic predicion. The main goal is o reduce burs assembly delay in Opical Burs Swiching (OBS) neworks. Research has shown ha raffic always change from ime o ime hence any measure ha is pu in place should be able adap o such changes. Wih our proposed burs assembly algorihm, he raffic rae is prediced and he prediced rae is used o dynamically adjus he burs assembly lengh. processing and configuring delay. This is advanageous in he sense ha by he ime he daa burs arrives a he node, he node would be already configured hus minimizing delays on he daa burs. There is however a problem of saisical muliplexing gain due o burs-level granulariy which is no he case wih OCS. One problem is encounered by OBS due o fac ha i uses saisical muliplexing conenions a he node. The focus of his paper is o explore burs assembly algorihms in opical burs swiching neworks and ulimaely propose an efficien burs assembly algorihm which uses ime and burs lengh wih raffic predicion. Figure shows an OBS node archiecure. Index erms Opical Burs Swiching, Burs assembly algorihm, Traffic predicion. INTRODUCTION Opical swiching neworks have been proposed for Nex Generaion Neworks (NGN). Various swiching paradigms of such neworks have been proposed. These include are Opical Circui Swiching (OCS), Opical Packe Swiching (OPS) and Opical Burs Swiching (OBS). All hese were discussed in []. However OBS has gained a lo of ineres hence curren sudy is focused on i. The concep of burs swiching was brough by he desire o overcome some of he problems associaed wih OPS and OCS such as he need for buffering and deflecion, synchronizaion and high conrol overheads when payloads are small. A comparison of hese hree swiching paradigms is shown in able. Unlike he oher paradigms which use a fixed packe size, in OBS uses a variable size packe. This packe is normally called a burs. The burs ransmis IP packes which have he same desinaion and oher common aribues hrough a bufferless nework. Thus, he burs will no go hrough O/E/O conversion a inermediae nodes. Each burs is associaed wih a conrol packe (CP) which conains header informaion of he packe and also he lengh of he burs. The CP is usually smaller han he daa burs. Daa burs and conrol packe are hen ransmied on separae channels. The daa burs and he CP are sen wih an offse ime beween hem. Only he conrol goes hrough O/E/O conversion a inermediae nodes. CP will be ahead and will be processed elecronically o configure he swiching fabric. The ime offse used o compensae for Figure : Archiecure of an OBS edge rouer [] This paper is presened as follows; Secion 2 gives an overview of burs assembly and basic burs assembly algorihms such as ime-based, burs lengh and mixed ime/burs lengh algorihm. Burs reservaion proocols are discussed in secion 3. As burs scheduling algorihms are also a very imporan aspec of OBS, hey are discussed in secion 4 whils secion 5 presens linear raffic predicion. In secion 6 a new burs assembly algorihm which is based on mixed ime/burs lengh wih raffic predicion is presened. Our simulaion scenario is presened in secion 7. This secion also gives he performance crieria used o compare he proposed algorihm agains he available algorihms. Simulaion, comparison and discussion of resuls are presened in secion 8. Secion 9 concludes his paper. Table : Opical Swiching paradigm comparison [2] Swiching Bandwidh Laency Proc/Sync Overhead Opical buffer Traffic adapabiliy Paradigm uilizaion Circui Low High Low No required Low Packe High Low High Required High Burs High Low Low No required High

2 2. BURST ASSEMBLY CONCEPT Burs assembly is a procedure whereby incoming packes from various sources are aggregaed ino burss a he edge of an OBS nework. As shown in figure, he swiching uni forwards he arriving packes o burs assembly unis. Packes which are bound for he same oupu lines are processed ino one burs assembly uni. A policy for assembling hese burss is herefore necessary. Some echniques which are used are ime hreshold based algorihms and burs lengh-based (size hreshold based) algorihms [3]. Anoher echnique which is here is he mixed ime/burs lengh algorihm. A brief review of hese echniques is given in he nex subsecion. 2. Time based algorihm Wih ime-based algorihms, a imer is sared a he sar of each burs assembly cycle. When a deermined fixed ime hreshold, T is reached, all packes ha would have arrived during ha would be assembled ino a burs. In his scheme a minimum burs size L min is se. Therefore all burss generaed should be of size equal o L min or more. If he burs size is less, hen is size is padded o size equal o L min. Careful consideraion should be made when choosing he value of T. Large values may lead o unnecessary packe delays a he edge. The oher problem is ha if he value is small, many small burss will be generaed and ha would lead o high conrol overhead. The algorihm is as follows [4]; Sep 0: While imer < T Accep new arriving IP packe Check imer. Sep : If burs_size < L min hen /*Check burs_size*/ pad i o L min Generae burs and send i oupu por 2.2 Burs lengh based algorihm Burs lengh based also known as size hreshold algorihm was proposed in [4]. Unlike he ime-base algorihm, his algorihm uses a fixed burs size L o decide he generaion of a burs. Once he L has been reached a burs is assembled and sen o he corresponding oupu por. The shorcoming of his scheme is ha depending on raffic, i may ake long for L o be reached and resuls in undesirable delays a he edge nodes. 2.3 Time/ lengh hybrid algorihm To deal wih he problems faced by he wo schemes previously discussed, a hybrid of he wo schemes was proposed in which a burs is assembled when eiher L or T is reached, whichever comes firs. When T is reached firs and he burs size is no equal or more han he se L min, hen he burs size is padded o L min. The algorihm is implemened as follows;[4] Sep 0: While imer < T Accep new arriving IP packe If burs_size L hen Generae burs and send i o oupu por Sep : /* means imer expires */ If burs_size < L min hen /* Check burs_size*/ pad i o L min Generae opical burs and send i o oupu por The main disadvanage of he above burs assembly algorihms is ha hey are no adapive o raffic. In secion 5 a burs assembly algorihm which is adapive o raffic by way of raffic predicion is proposed. Afer a burs has been formed resources such as available oupu por or free channel need o be reserved for i o be sen ou o he oupu. In ligh of his burs reservaion proocols are discussed in secion BURST RESERVATION PROTOCOLS There are quie a number of burs reservaion proocols for OBS. However here are hose which researchers have gained more ineres in. In his secion we discuss one-way and wo-way reservaion proocols. 3. One-way reservaion proocols In one-way reservaion proocols, when a reques is sen here is no acknowledgemen. An example of one-way reservaion proocol is Jus-Enough-Time (JET) which was proposed in [5]. In JET, he conrol packe is sen ahead of he burs wih a se offse ime,. This reduces delays since here will be no ime required for an acknowledgemen. This also leads o low uilizaion of he edge rouer s access link. Anoher one-way reservaion proocol is called Jus-In-Time (JIT) [6]. JIT uses he same principle as JET wih a sligh difference. JIT uses an acknowledgemen from only he firs cross-connec. Since boh JET and JIT (afer firs cross connec) do no use acknowledgemens, some burss migh be los if hey are sen and heir ransmission requiremens are no me, hence one-way reservaion proocols have a high blocking probabiliy. 3.2 Two-way reservaion proocols Unlike one-way reservaion proocols, wo-way reservaion proocols make use of acknowledgemen. The mos imporan aspec of his proocol is ha when a reques has no succeeded he burs will remain in memory unil he reques succeeds. An example of wo-way reservaion proocol is wavelengh-roued opical burs swiching (WR- OBS) [7]. 4. BURST SCHEDULING ALGORITHMS Burs scheduling algorihms are mechanisms used o decide on which wavelengh o assign an ougoing burs. If here is no wavelengh available he scheduler will decide which fibre delay line (FDL) o use. There are wo

3 caegories of burs scheduling algorihms. There are hose which are based on Horizon and hose based on JET reservaion proocol. The scheduling Horizon is defined as he laes ime a which he wavelengh is currenly scheduled o be used [3.4]. The horizon scheduler selecs he wavelengh whose scheduling horizons are smaller han he burs s arrival ime. The scheduler hen updaes he scheduling horizon o be equal o he ime when he burs is due o be compleed [8]. Horizon scheduling algorihm is also known as Laes Available Unscheduled Channel (LAUC) wihou void filling algorihm. The aim of his algorihm is o minimize voids which are creaed when making new reservaions. When here are no available wavelenghs for selecion by horizon scheduler, he burs will be dropped. The fac ha his algorihm does no make use of voids makes i easy o implemen. However his resuls in undesirable burs loss. To solve his problem, versions of LAUC wih void filling were proposed in [9] and [0] o make use of voids such as LAUC wih Void filling (LAUC - VF), Minimum - Saring Void (Min - SV), Minimum Ending Void (Min - EV) and bes fi. 5. LINEAR TRAFFIC PREDICTION The Time/ lengh hybrid algorihm work well when raffic is moderae. However, when raffic load is high he problem of creaing many burs becomes apparen. In our proposed algorihm we seek o dynamically adjus he burs lengh by linear predicion oncoming raffic. There are a number of linear predicor models. They are auoregressive (AR), moving-average (MA) and auoregressive moving-average (ARMA). Some predicors which can also be used for raffic predicion are MSE, FARIMA and Gaussian. Mos of hese predicors ignore whie noise; hence hey do no adequaely predic raffic []. I was on his backdrop ha a linear predicor wih dynamic error compensaion (L-PREDEC) was proposed in []. Suppose he mean of raffic daa rae (bis/s) is o be calculaed for inerval [, ; 2 ), [ 2, ; 3 ),, where i+ - i =d is he size of he observing window, hen a ime series of he raffic rae R, R 2,, R is obained []. Applying a linear predicor on he ime series and based on correlaions srucures of he ime series, a linear predicor P R +h predics he fuure value R +h of a ime series R () based on hisory of R, R 2,, R.. However in pracice only limied iems can be used from pas hisory, hence; P R n + h = m + ai R i+ m) i= ( () where m is he mean of he ime series R (). L-PREDEC predicor is made-up of wo pars. Tha is; he linear predicion par; LP X + (2) where LP is he linear predicor and; The error compensaion par; where f ec is a funcion o calculae he error compensaion from he las predicion error. Therefore combining (2) and (3) we ge; P R + + ec ( = LP + f R LP R ) (4) The mean of he posiive bias of L-PREDEC is; x 2 exp x 0 2 2πσ 2σ err err where x = f ec []. dx 6. PROPOSED BURST ASSEMBLY ALGORITHM The proposed algorihm uses L-PREDEC o raffic rae as shown by equaion (4). Le he prediced value be R pred. The prediced value will be used o dynamically adjus he burs size according o a se values. The scheme uses a se of seleced burs sizes, [l, l 2, l 3,..., l n ] and se of raffic daa raes se, [R, R2, R 3,..., R n ], where n is number of elemens in each se. The chosen value of burs size, l n, will hen be applied o he Time/Lengh hybrid algorihm as shown in 2.3. Figure 2 shows he proposed burs assembly algorihm which uses raffic predicion. Proposed burs assembly algorihm Sep 0: Predic raffic rae, R pred (L-PREDEC mehod) Sep : Adjus burs size /*according o R pred */ If R pred R hen L = l if R pred R && R 2 hen L = l 2 if R pred R n- && R n hen L = l n L = l max Sep 2: While imer < T Accep new arriving IP packe If burs_size L hen Generae and send burs o oupu Go o Sep 2 Go o Sep 2 Sep 3: /* means imer expired */ If burs_size < L min hen /* Check burs_size*/ pad i o L min Generae burs and send i o oupu por Figure 2: Proposed burs assembly algorihm (5) f ec (X -LP - X ) (3)

4 7. SIMULATION SCENARIO Prop_Bursy, Hybrid_Bursy and Time_TH_Bursy. Time hreshold based bursy assembly performs poorer han ohers in boh uniform and bursy raffic. While he hybrid algorihm ou-performs he proposed algorihms under uniform raffic, he proposed algorihm performs beer han hybrid algorihm in bursy raffic. Figure 3: Simulaion opology In his secion we invesigae he performance of he hree algorihms. To do his we divide raffic load ino hree classes, which are ligh, medium and heavy load. Each class will hen be associaed wih a burs size. This is done so ha each raffic load can be assigned an appropriae burs lengh. The value of he assembly ime hreshold, T will be kep consan and same for all algorihms and packe delay will be invesigaed. Performance is evaluaed under uniform and bursy raffic. Figure 4: Burs assembly delay vs Load A raffic generaor (Poisson Pareo generaing source) is used o generae random daa raes from 0Gbps o 60Gbps. The burs size se is; [25, 60, 00] in kbyes. The hreshold ime, T value used are; T =0.ms, T 2 = ms, T 3 =0ms and T 4 =5ms. L min = 5kbyes and L max = 00 kbyes. The packes delay is measured using NS-2 simulaor. For he proposed scheme, raffic classes are as follows; for ligh raffic daa rae 20Gbps, medium daa rae 20Gbps bu 40Gbps and heavy raffic daa rae 40Gbps. Burs size is adjused o 25kbyes for ligh raffic, 60kbyes for medium and 00kbyes for heavy raffic. Time _ hreshold = α * d (6) Lengh _ Threshold = E[ L ] (7) p where α is used o adjus he ime hreshold parameer. E L ] and d are he mean packe lengh and average [ p assembly ime respecively. Average assembly ime is given by; E[ L p ]* N d = (8) M * Arrival _ Rae where M and N are he number of cliens neworks and desinaion neworks respecively. 8. RESULTS The resuls as shown in figures 4 and 5 indicae ha all burs assembly algorihms performance is poor when raffic is bursy and beer in uniform raffic. In figure 4; proposed (Prop), Hybrid, Time_TH (Time hreshold) show resuls under uniform raffic. Resuls for bursy raffic are; Figure 5: packe delivery vs load In figure 6, he resuls show ha he proposed algorihm has low packe loss percenage. When he performance of he hree algorihms is measured by considering he number of packes delivered, he proposed algorihm performs beer han ohers and he Time hreshold based algorihm performs poorly under boh uniform and bursy raffic. The resuls in figure 7 shows he padding percenage of he ime based, hybrid based and he proposed burs assembly algorihms. The resuls show ha when raffic is low he padding percenage is high for all he hree algorihms bu he padding is less in he proposed algorihm. As raffic load increases, padding percenage is reduced considerably and i is observed ha he padding percenage for he proposed algorihm reduces quickly.

5 Figure 6: packe loss vs load [4] B. Kanarci, S. F. Okug, T. Amaca, Performance of OBS echniques under self-similar raffic based on various burs assembly echniques, Compuer Communicaions, 2007 pp. 35~325 [5] M. Yoo and C. Qiao, "Jus-Enough-Time (JET): A high speed proocol for bursy raffic in opical neworks", IEEE/LEOS Conf. on Technologies For a Global Informaion Infrasrucure, pp , Aug [6] J. Wei and R. McFarland, Jus-in-ime signaling for WDM opical burs swiching neworks, J.Lighwave Technol., vol. 8, no. 2, pp , Dec [7] P. Bayvel, Wavelengh rouing and opical burs swiching in he design of fuure opical nework archiecures,in Proceedings of ECOC, volume 4, pages 66-69, 200 [8] Y. Chen, J.S. Turner, Pu-Fan Mo, Opimal burs scheduling in opical burs swiched neworks, journal of lighwave echnology, vol. 25, no. 8, pp augus 2007 [9] Y. Xiong, M. Vandenhoue, and H. C. Cankaya, Conrol archiecure in opical burs-wiched WDM neworks, IEEE J. Sel. Areas Commun., vol. 8, no. 0, pp , Oc [0] J. Xu, C. Qiao, J. Li, and G. Xu, Efficien channel scheduling in opical burs swiched neworks, in Proc. IEEE INFOCOM, 2003, vol. 3, pp [] Y. Wei, J. Wang, C.Wang, A raffic predicion based bandwidh managemen algorihm of a fuure inerne archiecure, icinis, pp , Third Inernaional Conference on Inelligen Neworks and Inelligen Sysems, 200 [2] W. Cui, M. A. Bassiouni, Virual privae nework bandwidh managemen wih raffic predicion, Compuer Neworks 42 (2003) Mmoloki Mangwala received his Bachelors of Science degree in Compuer Engineering in 2007 from Universiy of KwaZulu Naal and has submied his hesis for Masers of Science degree a Norh Wes Universiy. His research ineress include Nex Generaion Neworks and e-governmen. 9. CONCLUSION Figure 7: Padding percenage vs load In conclusion, he resuls of he simulaion show ha he proposed algorihm performs beer han oher bu only has more delay han he hybrid algorihm under uniform raffic. However, hese performance merics alone do no give he verdic o say he proposed algorihm is beer han he oher algorihms. In he fuure i will be necessary o evaluae he effecs of he proposed algorihm on raffic selfsimilariy. 0. REFERENCES [] Y. Chen, C. Qiao, and X. Yu, Opical Burs Swiching (OBS): A new area in opical neworking research,ieee Nework Magazine. vol. 8, pp. 6-23, May 2004 [2] A.K Garg, Analysis of burs/packe assembly echniques in high-speed opical swiching nework, Opik-Inernaional Journal for Ligh and Elecron Opics, Vol, 22, Issue. 7, pp (20) [3] A. Ge, F. Callegai, L.S. Tamil, An opical burs swiching and self-similar raffic, IEEE Commun. Le. 4 (March 2000)