Modified Opportunistic Deficit Round Robin Scheduling for improved QOS in IEEE 802.16 WBA networks

(IJCSIS) International Journal o f Comput er Science and Info rmation Security, Vol. 6 , No. 2 , 2009 Modified Opportunistic Deficit Round Robin Scheduling for improved QOS in IEEE 802.16 WBA networks C.Kalyana Chakravarthy Dept. of CSE M.V.G.R.C ollege of Engine erin g Vizianagaram, India kch.chilukuri@gmail.com Dr. P.V.G.D.Prasad Reddy Dept. of CS&SE Andhra University Colleg e of Engineering Visakhapatnam , India prof.prasa dreddy@ gmail.com Abstract — Packet and flow scheduling algorithms for WiMAX has been a topic of inte rest for a long ti me since the ver y inception of WiMAX networks . WiMAX offers advantages particularly in terms of Quality of s ervice it o ffers over a longer range at the MAC level. In our work, we propose t wo credit based scheduling scheme s one in which completed flow s distributes the left over credits e qually to all higher priority uncompleted flows(ODRREDC) and another in which completed flows give away all the excess cred its to the high est priority uncompleted flow(ODRRSDC). Both the schemes are compatible with 802.16 MAC protocol and ca n effic iently serve real time bursty traffic with reduced late ncy and hence improved QOS for real time flows. We comp are the two propos ed schemes for th eir latency, bandwidth utili zation and throughput for real time bur st flows with the opportunity base d Deficit Round Robin scheduling scheme. While the ODRR scheduler focuses on reducing the credits for the flows with errors, ou r approach also distrib utes these remaining cr edits together with the credits from completed flows equally among the higher pr iority uncompleted flows or totally to the highest pr iority uncompleted flow. Keywords- component; sche duling; quality of service; latency ; I. I NTRODUCTION IEEE 802.16 in PMP mode, defines five types of sch eduling services[1] to support quality o f service. They can be classified as Unsolicited Gr ant Services(UGS), Real-ti me Polling services(rtPS), Extend ed rtPS, non Real-time polling services(nrtPS) and Best E ffort(BE). Application of Unsolicited gran t services (UGS) is Voice ov er IP (VoIP) without silence su ppression. The mandatory service flow pa rameters that define this service are maximum sustained traffic rate, maximum latency, to lerated jitter, and request/transmission policy. Applications of Real-time Polling serv ice (rtPS) are Streami ng audio and video, M PEG (Moti on Picture Expert s Group) encode d. The mandat ory service flow param eters that define this service are minim um reserved traffic rate, maximum sustained traffi c rate, m aximum l atency, and request/transmission policy. Application of Extended real-time is VoIP with silence suppr ession. The mandato ry service flo w parameters are guaranteed data rate a nd delay. Application of Non-real-time Polling service is File Transfer Protocol (FTP). The m andatory service fl ow parameters to define this ser vice are minimum reserved traffic rate, maximum sustained traffic rate, traffic priority, and request/transmission policy. Applications of Best-ef fort service (BE ) are Web browsi ng, data t ransfer. The m andatory service f low parameters to define this service are maximum sustained traffic rate, traffic priority, and request/transmission policy. In WiMAX, the MAC layer at the base station is fully responsible for allocating bandwidth to all us ers, in both the uplink and the downlink. The only ti me the MS has some control over bandwidth allocation is when it has multiple sessions or c onnections wit h the BS. In that case, t he BS allocates bandwidth to the MS in the aggreg ate, and it is up to the MS to apportion it among the multiple conn ections. All other sched uling on t he downli nk and uplink is done by the BS. For the downlink, the BS can allocate bandwidth to each MS, based on t he needs of the incom ing traffic, without involving the MS. For the uplin k, allocations have to be based on requests from the MS. Different co nnection m anagement strategies have been pr o pose d, but the m ost common one is of m anagement connections first, r eal-time connections fo llowed by no n-real time connections and finally Best Effort connectio ns. In our work, we propo se and compare two credit based scheduli ng schem es, Opportunistic Defic it Round R obin Scheduling with Equa l Distribution of Credits and Opportunist ic Deficit Round Robi n Scheduling with Single Distribution of Cr edits. The first on e based on distribu tion of excess credits equally between all h igher priority flows while the other pr oposed schem e is based on dist ribution of excess 75 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal o f Comput er Science and Info rmation Security, Vol. 6 , No. 2 , 2009 credits to the highest priority flow which is yet to be completed. The schem es are used to schedule flows between two classes of flows, real- time and non real-time flows. We compare the two schemes in terms of the QOS parame ters namely the throughp ut, bandwidth utilization , maximum late ncy etc., and observ e that though the former on e is based on fair schedu ling, the latter in fact offers better performance under sim ilar co nditions compared to the opportunit y-based DRR schedul ing schem e. II. P REVIOUS W ORK A significant am ount of work has already gone int o scheduli ng discipli nes that pr ovide dela y guarantees and fairness. Time stamp scheduler ess entially uses the idea of assigni ng time stamps to packets and then transmitting the packets in some order that achi eves fair ness.WFQ[3] and WF 2 Q[4] algorithms fall into this category. However, both of the schem es require a reference wit h the GPS se rver to be maintained. Variants of W FQ include Self-Clocked FairSchedulin g [5] and Virtu al Clock [6], whic h do not need to maintain a reference GPS se rver and hence can c o mpute the time stamp in a more efficient way. Though time stamp schedulers ha ve good delay properties, t heir processi ng time is quite high. Round-robin schedulers [ 7][8][9][10] are the othe r broad cl ass of work-c onserving sche dulers. These schedulers typically assign tim e slots to flows in some sort of round-robin fashion. Tho ugh they ha ve better com plexity compared to packet schedulers, however t hey have poor delay characteristics, particularly for packets of varying sizes. Several im provements have been proposed to improve the de lay properties of the b asic Round-robin scheduler. Ther e is another class of algorithms that try to combine th e tight delay bou nd of time stam p based schedulers a nd the low tim e complexit y of round robin base d schedulers. They us ually adopt a basic roun d robin like sched u ling policy plus time stamp based sc heduling on a reduced nu mber of units [11] . Bin Sort Fai r Queueing [12] i s based on arr anging packets int o different bins base d on their t ime stam ps and scheduling in a FIFO manner . Stratified R ound Ro bin [13 ] uses the round r obin approach for inter-class scheduling and the ti me stamp approach for intra-class sche du ling after gr ouping fl ows into respective classes. Recently proposed algorithms like ADRR [14] enhance the de ficit round robin sc heduling discipline by taking into account the channel quality expe rienced by the transmitting node. The ADRR sche duler is designe d to achieve performa nce isolation among links charact erized by heterogene ous channel conditions. In the DRR sc heme, Stocha ic fair queui ng is used t o assign fl ows to queues . For servici ng the que ues, Round-r obin servicing is used, with a qua ntum of se rvice attached to e ach queue. It di ffers from the traditional Round-robi n in that if a queue is unabl e to send a packet in the previous roun d because a packet was too large, the rem ainder from the pre vious quantum is added to the quantum for the next round. Q ueues that are not completely serviced in a round are compensated in the next ro und. Howeve r, once a flow is serviced, irres p ective of its weight, it must wait for N − 1 other flo w s to be serviced until it is serviced again. Al so, during each round, a fl ow transmits its entire quantum at once. As a result, DRR has poor delay and burstiness prop erties. The Smoothed Round Robi n discipline addresses the output b urstiness problem of DRR. This is done by spreadin g the quantum allocated t o a flow over an e ntire ro und usin g a Weight Spread Sequence. Althou gh SRR also results in better delay bounds th an DRR, the wors t case delay experienced by a packet is still proportional to N , the number of flows. The Opportunity-b ased Deficit Round Robin sched uling scheme [16] is an improvement over th e DRR scheme in that it considers the ch annel status in d ecisions it makes in serving flo ws. Opportunity -based scheduling ensures that a flow is not allowed to win a larger allocation of the resource if it uses its allocat ed resource inefficiently. ODRR used a penalty factor define d as PenaltyFactor = S I / S A to achieve fairness in allocation of resource S, where S I is the ideal number of bytes transmitted per unit of resource consumed and S A is the actual number of bytes transmitted per unit of reso urce S consum ed. With the use of the penal ty factor, during an exec ution of t h e O DRR scheduler, a fl ow that takes longer to tra nsmit a packet will have its d eficit counter decremented by a larger amount than anot her flow that takes less time to successfully transmit a packet. As in DRR, when the packet to be transmitted next is predicted, the deficit counter is decremented and if it falls below 0, the schedul er procee ds to serve the next flow after ad ding the req u ired quantum once it aborts transmission attempts from the current flow. The ODRR schedu ler removes a packet from the queue only if its transmission succeeds. Fig. 1. The ODRR Scheduler 76 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal o f Comput er Science and Info rmation Security, Vol. 6 , No. 2 , 2009 To understand how the OD RR scheduler works, consider the above scenari o in Fig. 1. Sinc e quantum size is 750, the packet with size 750 is pro cessed in the first round. In the beginning of seco nd round the cre dit count is incremented by remaining cred its 0 plus quantum size 750.So in the second round, actually packets with sizes 50,500 and 150 can be served. Howe ver because o f error in pac ket four, only 550 bytes can be served in the second round. Thus, the penalty factor = 550/700 =0.78571. Thus the DC b ecomes DC=750-.7 8571*700= 75 0-550 = 200 .Flow 2 is ad ded to the que ue. So when the flow 2 is scheduled ne xt, after ot her flows wi thout errors are complete(not shown in figure) , it receives only 200 + 750 credits, which is enou gh to send the packet number 4,t ogether with the other packets successfully. Clearly, the ODRR scheme is an improvemen t over the DRR schem e in terms of fai rness and the th roughput achie ved. However, the problem with ODRR is that if a flow becomes backlogged due to erro rs, it has to wait until all the other subsequent flows without errors are completed before it is scheduled again . This may result in a large delay, and is undesirable, particul arly if th e flow is a high priority one and has encounter ed an error in the be ginning of a round. The situation is similar to DRR[2] if all flows encounter errors in packet transmissions at the beginning each round. The above probl em could be co nsiderably allev iated if we allow for distribution of the unused credits. It has to be understood that t he distri bution of credits occurs in case if a flow is complete and it has balance cred its. Our scheme is diffe rent from ODRR in the following aspects: 1) It p rovid es for sharing o f credits among uncompleted flows, allowing them to complete faster. 2) In ad dition to reducing the cred its for a flow with error, we try to increase the credits for oth er uncompleted flows withou t errors enabling them to complete faster. 3) It does not consider th e penal ty factor when distributing credits.T his reduces the overhead of determini ng the fraction of the packets trans mitted with errors to those transmitted without erro r in case of an error, the balance credits are simply the quantum size minus the sum of credits of successfully transmitted pa ckets, excluding the c urrent packet with error. Our proposed algorithm effectively reduces the latenc y betwee n flows whi le at the sam e time providing an im proved throughput an d ensuring f airness. III. O PPURTUNISTIC D EFICIT ROU ND ROBIN SCHEDULING WITH EQUAL DISTRIBUTIO N OF CREDITS ( ODRREDC ) In our pr oposed algo rithm, w hen a flow e ncounters an error , it is suspe nded and adde d to a queue. All the other completed flows distribute their bala nce credits among the higher priority flows. Thus all the higher priority flows that are incomplete in a round receive some additiona l credits, in addition to the quantu m size in their next round. This allows the other flows without errors to complete faster, which in turn reduces th e delay for processing the flows with errors in the queues in the order of their priority. Our model uses Inter-class scheduling for servicing the fl ows. It assum es fixed schedulin g intervals bet ween flows associated with a particular flow class. For each class F k , the length of a sc heduling inte rval is always 2 k slots. If a scheduling i nterval fo r F k starts at slot t , the next scheduli ng interval for F k starts at slot t + 2 k , and so on. A flow is backlogged if it has not receive d it’s fair share of bandwidth, i.e it still requires to be serviced in the next rounds. Backlogged flows a re considered to be a ctive. After every pending flow is serviced in the cu rrent time slot, clock time is t c is incremented. Otherwise, t c is adva nced to the earliest time when some flow class becomes pending again. Also, in our m odel the bandwidth is shared e qually betwee n the flows . Fig. 2. The Simulation Setup The following scenario explai ns the operation of the ODRREDC. For simplicity, we have chosen the qu antum size to be at least equal to the m a ximum packet size and the service pointer advanc es after each flow ha s been serviced. Fig . 3. Beginning of Round on e 77 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal o f Comput er Science and Info rmation Security, Vol. 6 , No. 2 2009 Fig . 4. End of Round one Fig . 5. Beginning of Round two Fig . 6. End of Round two Fig . 7. Beginning of Round three Fig . 8. End of Round three Fig . 9. Beginning of Round four for ODRREDC At the beginning of ro und 4, the entire 650+75 0 credits go to the first flow in the queue with error, left over credits plus 750 go to sec ond flow in the que ue in the next(fi fth) round. It can be observ ed that if f low 1 was not supplied with the excess credits, it would have taken one more additional round to complete. It can be noted t hat flows w h ich are c omplete do nate their d ebit to the highest priority flows yet to be completed, while othe r flows proceed the sam e way as in the DRRscheme. Fig. 10. The ODRREDC algorithm 78 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal o f Comput er Science and Info rmation Security, Vol. 6 No. 2 , 2009 IV. OPP U RTUNISTIC DEFICIT ROUND ROBIN SCHEDULI NG WITH SINGLE D ISTRIBUTION O F CREDITS ( ODRRSDC) In our pr oposed algo rithm, w hen a flow e ncounters an error , it is suspende d, added to a que ue. All the other completed flows distribute their balance credits to the h ighest priority pe nding flow. Thus hi ghest priority flow s that are incomplete in a round receiv e some additional credits, in addition to the quantu m size in their next round. This allows the other flows without errors to complete faster, which in turn reduces th e delay for processing the flows with errors in the queues in the order of their priority. The SCBSS diffe rs from other sched uling schemes as in [15] where generally a completed flow distributes its credits continuousl y in subse quent rou nds to the hi gher prio rity flows until it has no more credits to distribute. ODRRSDC operates the same as ODRREDC till end of ro und three in Fig .8. However, at the be ginning of round four, all 50 credits from the com pleted flow are given away to the highest priority flow 1. T his is sh own in t he figure bel ow. Fig. 11. Beginning of Round four fo r ODRRSDC Fig. 12. The ODRRSDC algorithm V. PROOFS Theorem I: The proposed algorithm ODRREDC is better or atleast equa l in terms of throughput and fairness co mpared to the ODRR. Lemma 1:In an exec ution o f the ODR R sche duler, at the en d of r ound k, 0 ≤ DC i (k) ≤ M for any fl ow i, where M is the maximum packet size. The lemma is identical to that in the case of DRR and the proo f can be found in [2] . Lemma 2: Dur ing an exec ution of t he ODRR sche duler over any m round s, for any flo w i , mQ i – M ≤ SPT i (M) ≤ mQ i + M where SPT i (M) is the total potential th roughput that can be achieved by a flow in m rounds, M is the maxim u m packet size and Q i is the quantum size to be added to a flow before the starting of each roun d. The potential throughput that a flow i may achieve during a ro und K is PT i (K) = Q i + DC i (k-1) - DC i (k) m SPT i (M) = ∑ PT i (K) = mQ i + DC i (0) - DC i (m) -1 k= 1 Appl ying lem ma 1, the statement of t he lemm a can be prov ed. For our pr oposed m ethod, m SPT i ’ (M) = ∑ PT i ‘ (K) = mQ i ‘ + DC i ’ (0) - DC i ’ (m) -2 K=1 Since in the propose d method , we are not re ducing the quantum size given to packets with de ficit credit s, Q i ‘ ≥ Q i for all m - 3 and since all completed flows distribute their balan ce credits to hi gher priority flows ,for any round there ex ist flows with DC i ’ (m) ≤ DC i (m) . -4 From equati on 1-4, it can be deduce d that SPT i ’ (M) ≥ SPT i (M) for som e m Theorem II: The proposed algo rithm provides fairne ss atleast eq ua l to that of the ODRR. The fairness m easure based on potential throughput measured acr oss interval (t 1,t2) is given by FM( t 1 ,t 2 ) = max V(i,j) [ SPT i (t1,t2)/w i - SPT j (t1,t2)/w j ] Wher e w i and w j are the weights of flows i and j respectively. For our proposed method, FM ’ ( t 1 ,t 2 ) = max V(i,j) [ SPT i ’ (t1,t2)/w i - SPT j ‘ (t1,t2)/w j ] It has already been proved that SPT i (t1,t2)/w i - SPT j (t1,t2)/w j ≤ Q + 2M from lemm a 2. Since for our prop osed method Q = Q + δ , It follows that SPT i (t1,t2)/w i - SPT j (t1,t2)/w j ≤ SPT i ’ (t1,t2)/w i - SPT j ‘ (t1,t2)/w j 79 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal o f Comput er Science and Info rmation Security, Vol. 6 , No. 2 , 2009 Ö FM(t1, t2) ≤ FM ’ (t1, t2) for a n y interval (t1,t2) which completes our pro of. VI. SIMULATION RESULTS We use a custom simulato r written in java. The simu lation runs in two thread s - the flow generator that generates packet an d the sche duler t hat checks at every configurable s cheduling period and sc hedules the packets. Both these modules ca n be ru n either con currently or independent ly. Sim ulation ha s been carried out on 20 queues, each containing maxim u m pack ets of variable size, for different quant um sizes for 20 seco nds an d the resul ts have been evaluate d. The pac kets are ge nerated acc ording to Poisson arrival pro cess. For our results we limited the number of flows s o that the sum tot al of thei r minim u m bandwidt h requirements matches the maximum capacity of th e network. We have chosen a 9KbpS outp ut queue and each input queue has six packet s of maxim u m size 750 by tes and has a bandwidth of 4500 bps. All flow s are critical and are arrang ed in the decreasing order of their priorities. Our algorit hm has shown reasonable im provement in terms of latency of critical flows, which makes it suitab le for real time comm unications such as real time Video- on demand. If all latency critical flows meet the requirements, the maximum delay between laten cy critical flows should not exceed (n * s) + Max/B where n is number of latency critical flows, B ba ndwidth o f the output line, s is m aximum size of packet in a fl ow, Max is m aximum quantum size. Fig. 13. Flow ID vs bandwidth utilization Fig 14. Flow ID vs Avera ge latency VII. CONCLUSIONS In our work, we have proposed two scheduling schemes ODRREDC and ODRRSDC for scheduling real time flows. It was observed from the results that while both the schemes perform better compared with the Opportunistic Deficit Round Robin scheduling scheme, the former is more suitable for real time flows under unsteady traffic conditions. In our method, any excessive idle bandwidth is reallocated to avoid wasting of available transmission capacity. In both cases, we assume scheduling under erroneous channel conditions. Scheduling on Multiple Input Multiple Output channels with m u ltiple antennas, scheduling on multi-hop networks for end to end service guarantees are also areas that demand further improvement. R EFERENCES [1] A. Sayenko, O. Alanen, J. Karhula and T. Hämäläinen,” Ensuring the QoS Requirements in 802.16 Scheduling”, Proceedings of IEEE/ACM MSWiM 2006, Torremolinos, Spain, Oct. 2006. [2] M. Shreedhar and G.Varghese,”Effici ent fair queuing using deficit Round Robin”, in Proc. SIDCOMM ’95 Boston, MA, Aug 1995. [3] Demers,A.S Keshavand S. Shenkar,1989, ”Analysis and sim ulation of a fair queuing problem”, Pro ceedings of the Symposium and Comm unications Architectures and protocols, September 25-27,A.C. M NewYork, USA pp:1- 12. [4] Bennet , J.C.R and H.Zhang,1996, ”WF2 Q: Worst –case fair weighte d fair queuing”, Proceedings of the INFO COM, March 24-28,San Fransisco,CA, pp:1-9 [5] Golestani,S.J.1994,”A self clocke d fair queuing schem e for broad band applications”, Proceedings of the 13 th IEEE INFOCOM’94, Networking for Global Communications, June 12-16, T o ronto, Ont, Canada, pp: 636-646. [6] L. Zhang, “A new architectu re fo r Packet switched network pr otocols”, PhD dissertation, Massachesets Institute of technology, July 1989 [7] Lenzini, L., Mingozzi, E., and Stea, G. Aliquem: “a novel DRR impl ementation to achieve better latency and fairness at O(1) complexity,” In IWQoS’02 (2002). [8] “The Sm oothed Round-Robin Sc hedul er”, Paul Souther ington, Member, IEEE, ECE742, 28 APRIL 2005. [9] B. Bensaou, K. Chan, and D. Tsa ng, “Cre dit-based fair queuing (CBFQ): A simple and feasible scheduling algorithm for packet networks”, IEEE ATM97 Workshop, pp. 589594, May 1997. [10] Dessislava Nikolova and Chris Blondia, “Last-Backlogged Fir st-Served Deficit Round Robin (L BFS-DRR) Packet Scheduling Algorithm”, 15 th IEEE International conference on networks,Nov.2007. [11] Deng Pan, Yuanyuan Yang, “Cred it Based Fair Scheduling for Packet Switched Networks”, IEEE INFOCOM’05. [12] S. Cheung and C. Pencea, “B SFQ: bin sort fair queuing,” IEEE INFOCOM ’02, pp. 1640-1649, New York, Jun. 2002. [13] S. Ramabhadran, J. Pasquale, “Str atified r ound robin: a low complexity packet scheduler with bandwidth fairness and bounded delay,” ACM SIGCOMM ’03, pp. 239-250, Karlsruhe, Germany, Aug. 2003. [14] Riggio, R. Miorandi, D. Chlam tac, “ Airtime Deficit Round Robin ( ADRR) packet scheduling algorith m,” 5 IEEE International conference on Mobile Ad Hoc and Sensor Systems-MASS’08. th 80 http://sites.google.com/site/ijcsis/ ISSN 1947-5500 (IJCSIS) International Journal o f Comput er Science and Info rmation Security, Vol. 6 No. 2 , 2009 AUTHORS P ROFILE [15] Tsung-Yu Tsai, Z sehong Tsai, “Des ign of a Packet Scheduling Schem e for Downlink Channel in IEEE 802. 16 BWA Systems”, in WCNC’08 (2008). C.K alyana Chakravart hy has a teachin g experience of over nine [16] Yunkai Z hou, Madhusudan Hosaagrahar a and Harish Sethu,” Opportunity-Based Deficit Round Robin: A Novel Packet Scheduling Str ategy for Wireless Networks”, in Proceedings of the IEEE Workshop on High Performance Switching and Routi ng Kobe, Japan, May 26–29, 2002 Yea rs and is currently work ing as Associate prof. in M.V.G.R. College of En gineering, Vizian agaram. He has been activ ely working o n diverse areas of network cach ing MA NETs routing protoco ls, resource allo cation a nd sched uling in WiMA X, Mesh networks . Dr. P.V.G .D Prasad Reddy has a tea ching exp erience of over twenty years. He is currently serving as th e Registrar at the Andhra University. He has over 16 publications in Intern ational Journal s and 20 papers in conferen ces. His Research area s include Sof t Computing, Software Architectu res, Knowledge Disco very fro m Databases , Image Processin g , Number theo ry & Cryptosyste ms. 81 http://sites.google.com/site/ijcsis/ ISSN 1947-5500