On the Prediction of EPON Traffic Using Polynomial Fitting in Optical Network Units

Transcription

1 On the Prediction of EP Trffic Using Polynomil Fitting in Opticl Networ Units I. Mmounis (1),(3), K. Yinnopoulos (2), G. Ppdimitriou (4), E. Vrvrigos (1),(3) (1) Computer Technology Institute nd Press "Diophntus", Greece, (2) University of Peloponnese, Deprtment of Telecommunictions Science nd Technology, Greece, (3) University of Ptrs, Computer Engineering nd Informtics Deprtment, Greece, (4) Aristotle University of Thesslonii, Deprtment of Informtics, Greece, Keywords: Abstrct: Prediction, Ethernet pssive opticl networ, Polynomil prediction, Dynmic bndwidth lloction, Dely We propose trffic prediction lgorithm tht reduces the pcets dely in Ethernet Pssive Opticl Networs (EPs). The lgorithm relies on Multi-Point Control Protocol (MPCP) messge nd trffic monitoring t the Opticl Networ Units (Us) nd utilizes the monitoring informtion to predict the ccumulted burst size using higher order lest-men-squre polynomil pproximtions. The simultion of the lgorithm shows tht it chieves dely improvement of over 30% without ny further modifiction in the communiction nd bndwidth ssignment procedure of the EP. 1. INTRODUCTI Pssive opticl networs (Ps) [1],[2] re n ttrctive solution for the deployment of nextgenertion ccess networs, due to their low implementtion cost, simple opertion nd high-line rtes mde possible by the cpcity of opticl fibers. Ethernet Pssive Opticl Networs (EPs) in prticulr, which represent the convergence of lowcost Ethernet equipment nd low-cost fiber infrstructure, find widespred ppliction in locl nd metro re networs, supporting the fiber infrstructure tht is being instlled within the scope of fiber-to-the-home, building nd curb (FTTC) end-user ccess. In n EP networ, multiple opticl networ units (Us) ccess the shred chnnel to rech the opticl line terminl (OLT) through pssive opticl splitter. To rbitrte the multiple U ccesses, n effective bndwidth lloction scheme is required. The interleved polling scheme with dptive cycle time (IPACT) [3], [4] is implemented t the OLT. IPACT periodiclly receives bndwidth requests from ll connected Us nd llocte trnsmission slots ccordingly. The verge cycle time in IPACT contributes to the P system ltency, since Us re served in round-robin fshion nd ech U must wit for the full cycle durtion before being served gin. Thus, the verge cycle time nd consequently the dely, depends on the bndwidth lloction scheme implemented by IPACT. In generl, bndwidth lloction schemes cn be ctegorized s fixed or dynmic. Fixed bndwidth lloction (FBA) schemes [5] utilize equl size time-slots nd offer fixed time-slot to ech U irrespective of its trffic lod. The U-to-OLT (upstrem) communiction chnnel is therefore reserved even when the ctul U trffic is not sufficient to fully utilize the slot nd this bndwidth underutiliztion leds to trnsmission gps nd incresed frme service times. On the other hnd, dynmic bndwidth lloction (DBA) [6], [7] ssigns the bndwidth in n dptive fshion bsed on the current trffic lod of ech U. The ide in DBA schemes, such s the one implemented in IPACT, is to re-distribute bndwidth from lightlod to hevy-lod Us within single cycle durtion nd consequently fully utilize the vilble cpcity, thus reducing the overll P ltency. An improvement on the P ltency cn be obtined by mens of trffic prediction, technique tht hs been widely studied in both wireline nd

2 wireless networs [14]. During the time of bndwidth negotition, nd in prticulr during the intervl tht lsts from the moment the U sends the bndwidth requirement until the moment it cn strt sending the buffered dt. More dt will rrive t the U buffer tht time nd will remin in the buffer until the next cycle. As result these will not be ten into ccount when the bndwidth request ws sent. OLT-bsed trffic prediction relies on estimting future "on-verge" bndwidth requirements for ll Us in the networ bsed on their previous bndwidth requests. A ey drwbc of OLT-bsed prediction, however, is tht it my not ccurtely identify, nd therefore respond, to rpid chnges in the U trffic. On the other hnd U-bsed prediction, cn be performed within single cycle, since Us re ble to constntly monitor incoming trffic, nd therefore cn dpt to trffic chnges significntly fster. Predictive techniques estblish mthemticl model tht processes the series of dt pcets in order to estimte the future trffic flow. A lrge vriety of trffic prediction lgorithms for EPs hve been proposed in the lst yers, in order to improve the bndwidth lloction strtegy nd the totl system performnce [7], [8]. These prediction techniques cn be executed t the side of OLT [9], [10] or the Us [4],[11],[12], with the pros nd cons of ech pproch tht hve been described erlier. The technique proposed in [9] consists of two-stge bndwidth request scheme. In the first stge, DBA is performed for the next cycle t the U level ssigning bndwidth to the Us tht hve more unstble (difficult to predict) trffic. In this wy it becomes esier to reduce the prediction error by shortening their witing times. In the next stge, liner prediction-bsed excess bndwidth request is done for the more stble Us. At the OLT, the proportionlly vilble bndwidth for n U is llocted to relted trffic clsses, strictly bsed on their respective requests ordered by their priority. In [10], the uthors propose prediction process tht is bsed on genetic expression progrmming to reduce the queue size vrition nd the pcet dely. Ting different pproch [4], [11], [12] propose prediction techniques tht re pplied t the Us. In [4], limited shring with trffic prediction scheme ws proposed nd shown to enhnce DBA process. For U-bsed trffic prediction nother pproch ws presented in [11] where uthors propose liner clss-bsed prediction model tht tries to estimte the incoming trffic until the next polling cycle. This model uses informtion from previous bndwidth requests in order to predict bndwidth request t ech U in the networ, ccording to the OLT priority clsses. The effect of long-rnge dependence of internet trffic in the prediction ws studied in [12]. While the prior wors hve used complicted prediction techniques t the Us, the estimtes they produce refer to single prmeter, tht is the bndwidth to be llocted, which is however complex metric (rtio of dt size over time durtion). Within the context of U-bsed trffic prediction, we propose novel lgorithm for decresing ltency in EPs. Our lgorithm () pproximtes the frme rrivls within the durtion of single EP cycle using lest-men-squre polynomils nd (b) estimtes the durtion of the upcoming cycle vi lest-mens-squres dptive filter. Subsequently, the two quntities re combined to produce the mount of dt tht the U will hve ccumulted by the time the next bndwidth ssignment from the OLT (GATE messge) rrives. The U then communictes the predicted rther thn the ctul dt to the OLT in the REPORT messge), thus providing the DBA mechnism with more informed guess of its trffic requirements. We show vi simultion tht the incorportion of the proposed prediction methods in the EP opertion cn reduce the frme dely from 25% up to 30% when compred to the stndrd opertion of the limited nd gted versions of Interleved Polling with Adptive Cycle Time (IPACT), depending on the trffic lod nd the burstiness of the incoming trffic. Moreover, this significnt performnce benefit is obtined by pplying the prediction lgorithms loclly t the Us nd without ny further modifiction on the Multi-Point Control Protocol (MPCP) procedures or the opertion of IPACT. At the sme time the proposed solution exhibits low computtionl complexity, which is prticulrly ppeling feture when considering the U processing cpbilities nd ssocited cost. The rest of the pper is structured s follows: Section 2 presents our proposed trffic prediction technique nd its scope of ppliction in EPs. Section 3 detils the simultion setup tht ws utilized to evlute the performnce of the prediction method. Section 4 discusses the results tht hve been obtined in terms of ltency. Finlly, Section 5 concludes the min contributions of this pper.

3 2. PREDICTI ALGORITHM BASED U In the stndrd EP opertion, the communiction between the OLT nd the Us tes plce by mens of n interleved polling scheme with vrible cycle time (IPACT). IPACT opertes in successive cycles, nd during ech respective cycle the OLT sends GATE messges tht crry bndwidth grnts to ll Us in the EP. The Us respond to the GATE messges nd send their dt in co-ordinted fshion, s specified in the GATE messges, so s to chieve collision free trnsmissions in the upstrem direction. In ddition to their dt, the Us lso inform the OLT bout their bndwidth requirements (buffer sizes) vi REPORT messges nd the IPACT cycle ends upon the reception of the REPORT messges from ll Us in the EP. At tht time, the OLT executes dynmic bndwidth lloction (DBA) lgorithm to clculte the grnts of the next cycle, nd new exchnge of GATE nd REPORT messges ensues. As result, the DBA does not te into ccount () dt tht hve been ccumulted t Us tht re served ner the beginning of the cycle nd re forced to report erly, or (b) dt tht will be ccumulted t Us tht re served towrds the end of the upcoming cycle nd will receive lte grnt. This leds to n dditionl dely of cycle time, which cn be prticulrly significnt especilly in IPACT vritions with incresed or infinite mximum cycle durtions. The dditionl dely cn be reduced in strightforwrd mnner by hving ech U perform prediction exctly before the genertion of the current REPORT messge by estimting its buffer occupncy for the instnt it will receive the next GATE messge. The U cn then use the REPORT messge to communicte the prediction to the OLT rther thn the ctul (current) buffer size. Our proposed prediction lgorithm of the U buffer size cn be summrized s follows: Step 1: Constntly monitor the incoming trffic from hosts in log file until GATE messge hs been received from the OLT. Step 2: Upon the reception of the GATE messge eep record of its rrivl time T(n-1). Step 3: Utilize the trffic log to estimte the instntneous buffer size B(t). Step 4: Utilize the rrivl times of previous GATE messges to predict the rrivl time of the next GATE messge T(n). Step 4: Combine B(t) nd T(n) to clculte the expected buffer size B(n) t the reception of the next GATE messge. Step 4: Trnsmit the llocted number of frmes in the received GATE nd then issue REPORT messge tht crries the bndwidth request B(n). Step 5: Reset the trffic log to the remining buffer size nd re-strt from Step 1. The presented lgorithm requires the estimtion of two ey prmeters: () the instntneous U buffer size B(t), nd (b) the rrivl time of the next GATE messge T(n+1). The estimtion of the instntneous buffer size is performed by monitoring the incoming frmes tht rrive between REPORT messges. To this end, the U cretes log of the frme size S i nd the rrivl time t i for ech frme tht is received. Ech frme rrivl corresponds to n increse of the number of bytes B i tht re stored t the U buffer, following: B B S, (1) i i 1 i while the remining queue size B 0 fter the U trnsmission t t 0 is used to initilize Eq. (1). Given (1), th degree polynomil eqution tht correltes the buffer size B(t) nd the elpsed time t is cn be clculted by the (t i, B i ) pirs, ccording to: 0 2 B t t t t t (2) where the coefficients 0, 1,, in the bove polynomil re clculted in lest-men-squres fshion by: 1... t1 0 B t1 1 B t n Bn Or T*A B (3) The U is ble to predict its queue sttus t ny given future time t nd up to the next GATE messge. The exct rrivl time of the next GATE messge, however, is not nown when the U cretes the REPORT messge nd s result the U hs to estimte it, s well. To this end, the U monitors the rrivl times of GATE messges nd predicts the rrivl time of the next GATE T(n) by mens of normlized lest-men-squre (NLMS) prediction filter which is given by:

4 Tble 1 Simultion Prmeters Physicl Lyer Prmeters IPACT Prmeters Trffic Prmeters Symbol Description Vlue (Limited - IPACT) N U Number of U s 8 N host Number of U Hosts 15 d U distnce 10 m R d R u Downstrem Line Rte Upstrem Line Rte 10 Gb/s 1 Gb/s R n Host Line Rte 100 Mb/s T mx Mx Cycle Time 2 ms Unlimited W mx Mximum Grnt Size bytes Unlimited (, ) Preto prmeter 1.2, 1.5, 1.8 b b 1.2 0, , , b (b, b ) Preto prmeter 1.5 0, , , , ,0016-0,00048 Prediction Prmeters p NLMS order 25 M NLMS step size constnt 0,0001 p n w i T n i Tˆ (4) i1 n where p is the filter order nd w n (i) re the filter co-efficients tht re updted t every cycle. 1 T n i wn i wn i M e n 1, i 1,, p, p 2 T n ˆ e n 1 T n 1 T n 1, 1 (5) the NLMS step size M hs constnt numeric vlue (Tble 1). 3. SIMULATI SETUP The performnce of our proposed lgorithm ws verified vi simultion experiments using the OMNET++ open source simultor [13]. In our setup, stndrd EP rchitecture interconnected n OLT with eight Us t distnces of 10 m, while the EP rtes were considered symmetric (10 Gb/s downstrem - 1 Gb/s upstrem). MPCP protocol forms type of mster-slve REPORT/GATE mechnism, which mens tht requirements re put forwrd by ech U nd re rbitrted by the opticl line terminl (OLT). The communiction model ws bsed on existing OMNET++ models tht provide the bsic MPCP functionlities t the OLT nd Us. Two IPACT lloction schemes were implemented t the OLT, nmely, the limited nd the gted version. For the limited-ipact implementtion, OLT grnts n upper bounded trnsmission window size per U. On the other hnd, in cse of gted- IPACT, OLT lloctes the estimted requested bndwidth for ech U in our networ. The incoming trffic for the purpose of the simultion ws fed to ech U from n opticl switch tht ggregted frmes from fifteen independent hosts (sources) Figure 1. The hosts trnsmitted dt in the form of fixed size 1000 byte Ethernet frmes t

5 4. RESULTS We hve conducted two sets of simultions experiments. The first set compres the limited IPACT lgorithm without prediction to our Figure 1: Simultion model line rte R n of 100 Mb/s. Ech of the hosts generted dt frmes independently of ech other, ccording to n / trffic model. This trffic model lso nown s Preto distribution consist of two different periods for ech host, n (busy) nd n (idle) period. Due to the form of the Preto distribution, (busy) periods were lwys followed by (idle) periods. The mthemticl formul of the distribution is described in Eq. 6: b f x (6) 1 x Figure 2: Men Dely for low burstiness trffic Prmeters nd b relte to the verge busy time durtions nd idle time durtions, T nd T respectively (eq. 7). T T b 1, b 1. (7) The vlues tht were used in our simultions for the prmeters, b nd, b of the nd periods, respectively, re presented in Tble 1. These vlues resulted in - periods with durtions t the msec time scle, which corresponds to single IPACT cycle, since n ccess-oriented P is not expected to remin idle for severl successive IPACT cycles. Given the bove verge busy nd idle periods of ech host, it ws possible to clculte the offered lods ρ in the P from the number of Us (N U ), the number of hosts per U (N host ) nd the individul host lod (ρ host ) s : Figure 3: Men Dely for mediun burstiness trffic NU Nhost host NU Nhost T T T (8) Figure 4: Men Dely for high burstiness trffic

6 corresponding scheme tht uses prediction lgorithms. The second set of simultions evlutes the performnce of gted-ipact lgorithm without prediction ginst our prediction lgorithms. For the purposes of the simultion, three different trffic burstiness scenrios tht correspond to low, medium burst nd high burst ( = 1.8, 1.5, 1.2, respectively) were used. Moreover, we evluted the prediction lgorithm for polynomils of degree equl to one (i.e., liner prediction) nd two, since higher degree for the polynomils led to severe prediction inccurcies tht negtively ffected the EP performnce. The respective results re shown in Figure 2-4 for the limited- IPACT nd Figure 5-7 for the gted-ipact. The results clerly demonstrte tht the Limited- IPACT performs in superior fshion when prediction bsed reports re sent by the Us. A percentile dely reduction of over 25% is observed for medium offered lods round 0.6, while smller benefit is observed s the lod becomes lighter. For higher lods, prediction only hs minor beneficil impct when the trffic is reltively smooth (=1.8 nd 1.5). As the trffic becomes significntly bursty (=1.2), the proposed liner prediction lgorithm cn be detrimentl in terms of ltency, minly becuse the cycle durtions become irregulr nd the GRANT rrivl times re not correctly clculted by the NLMS. As result, Us request the lrgest possible grnt nd IPACT performs in TDMA mnner with mximum durtion bndwidth grnts. For qudrtic prediction the dely results improve in ll cses, even for highly bursty trffic profile (Figure 4) nd the dely reduction is improved by up to 30% for medium offered lods round 0.6. As the lod increses, the prediction benefit reduces to under 10%; still, it is importnt to notice tht qudrtic prediction tends to correct the detrimentl effect of liner prediction with incresing burstiness. A similr behvior is observed for gted-ipact in Figure 5-7; the proposed liner prediction mechnism improves the verge dely in this IPACT vrition by 25% for medium lods s shown in the simultion results. An importnt difference with limited-ipact, however, is becoming evident for bursty trffic (=1.2) nd t hevy lods; in this regime even more extended bndwidth grnts re requested by the Us nd re llowed by the OLT, due to the fct tht gted- IPACT does not pose n upper limit on the size of the grnts. As result, the verge dely is lso incresed by significnt fctor. An even better performnce for the cse of gted-ipact is Figure 5 : Men Dely for low burstiness trffic Figure 6 : Men Dely for medium burstiness trffic Figure 7: Men Dely for high burstiness trffic observed for qudrtic prediction. As it cn be seen

7 from the dely results, n improvement of 26% cn be chieved in medium offered lods from 0.5 to 0.7 for ll degrees of trffic burstiness. Moreover, when the offered lods increse, the utiliztion of second order polynomils provides better dely performnce from its liner counterprt. Especilly for medium burst trffic (=1.5) the dely is ble to chieve profits up to 27%, round 0.8. Finlly, in ccordnce with the limited-ipact results qudrtic prediction lgorithm exhibits better stbility t high lods. 5. CCLUSI We presented n U bsed prediction method tht is pplicble in EPs. The method relies on the ppliction of polynomil fitting nd the Normlized Lest Men Squre (NLMS) lgorithms for the estimtion of the instntneous U lod nd IPACT cycle durtion, respectively, to predict the U buffer size t the time of its next trnsmission. We showed vi simultions tht if the predicted (estimted) buffer size, rther thn the ctul size, is reported to the OLT then significnt (over 25%) verge dely reduction cn be relized over stndrd EP when liner bsed prediction lgorithm is used. Also when the prediction method uses second order polynomil (nonliner prediction lgorithm) the verge dely improvement is over 30% for ll degrees of trffic burstiness. Moreover, the proposed techniques re totlly comptible with the bndwidth reporting nd lloction mechnisms tht hve been stndrdized in EPs, s well s with other populr well- IPACT vritions (Limited nd Gted). ACKNOWLEDGEMENTS This wor hs been funded by the NSRF ( ) Synergsi-II/EPAN-II Progrm "Asymmetric Pssive Opticl Networ for xdsl nd FTTH Access," Generl Secretrit for Reserch nd Technology, Ministry of Eduction, Religious Affirs, Culture nd Sports (contrct no. 09SYN ). [1] B. Muherjee, Opticl WDM Networs, Springer, University of Cliforni, Dvis. [2] Jue, Json P., Vorne, Vinod M., Opticl Burst Switched Networs, Opticl Networs Series, Springer. [3] G. Krmer, B. Muherjee, G. Pesvento, "IPACT: dynmic protocol for Ethernet P (EP)", IEEE Communictions Mgzine, vol. 40, no. 2, pp [4] Y. Luo, N. Ansri, "Limited shring with trffic prediction for dynmic bndwidth lloction nd QoS provisioning over EPs," OSA J. Opticl Networing, vol. 4, no. 9, pp [5] G. Krmer, B. Muherjee, nd G. Pesvento, Ethernet P (ep): Design nd Anlysis of n Opticl Access Networ, Photonic Networ Communictions, vol. 3, no. 3, pp [6] G. Krmer, B. Muherjee, nd A. Mislos, Ethernet Pssive Opticl Networs, In S. Dixit (Ed.), Multiprotocol over DWDM: Building the Next Genertion Opticl Internet, John Wiley & Sons, pp [7] M. Mcgrry, M. Reisslein, M. Mier, "Ethernet pssive opticl networ rchitectures nd dynmic lloction lgorithms," IEEE Communictions Surveys, vol. 3, no. 10, pp [8] N. Sde, A. Khotnzd, A dynmic bndwidth lloction using two-stge fuzzy neurl networ bsed trffic predictor, Proccedings of IEEE Interntionl Conference on Neurl Networs,Hungry, pp [9] I.-S. Hwng, Z.-D. Shyu, L.-Y. Ke, C.-C. Chng, A novel erly DBA mechnism with prediction-bsed fir excessive bndwidth lloction scheme in EP, Elsevier Computer Communictions, vol. 31, pp [10] I.-S. Hwng, J-Y Lee, A. Liem, QoS-bsed Genetic Expression Progrmming Predicition Scheme in the EP s, Progress In Electromegnetics Reserch Symposium Proceedings, 1589 [11] D. Swdes, S. Vibhv, G. M. Hri, S. Nvrti, R. Abhishe, A new predictive dynmic priority scheduling in Ethernet pssive opticl networs (EPs), Opticl Switching nd Networing, vol. 7, pp [12] D. Morto, J. Accil, L.A Diez, M. Izl, E. Mgn, On liner prediction of Internet trffic for pcet nd burst switching networs, IEEE ICCN [13] Omnet++ simultor, [14] Z. Zhu, Design of Energy-Sving Algorithms for Hybrid Fiber Coxil Networs Bsed on the DOCSIS 3.0 Stndrd, IEEE/OSA Journl of Opticl Communictions nd Networing, vol. 4, pp REFERENCES