A comparison of scheduling algorithms in HSDPA

Transcription

1 A omparison of sheduling algorithms in HSDPA Stefan M. Sriba and Fambirai Takawira Shool of Eletrial, Eletroni and Computer Engineering, University of KwaZulu-Natal King George V Avenue, Durban, 404, South Afria sribas@gmail.om or ftakaw@ukzn.a.za Tel: Abstrat This paper ompares the behaviour of Earliest Deadline First () and Opportunisti- (O-) with that of Round Robin, Proportional Fair Throughput (PFT), and Maximum Carrier to Interferene ratio (MaxC/I). Video, voie and web traffi are transmitted through an HSDPA air interfae. The resultant queueing delay and fairness are measured and ompared for eah sheduler. Index Terms Earliest Deadline First (), queueing delay. I. INTRODUCTION As part of the UMTS standard, the third generation partnership projet (3GPP) inluded High Speed Downlink Paket Aess (HSDPA) in Release 5. In HSDPA, all resoures are usually made available to a single user on a Transmission Time Interval (TTI) basis []. HSDPA distinguishes itself from the previous WCDMA standard, by using an Adaptive Modulation and Coding (AMC) sheme, whih enables it to respond rapidly to hannel flutuations, without the need for fast power ontrol, whih is disabled in HSDPA. Furthermore, Variable Spreading Fators have also been disabled, while the TTI has been dereased from 0ms to 2ms, allowing for muh faster sheduling responses. Finally, the model makes use of a Fast Physial Layer Hybrid ARQ sheme. This paper ompares the behaviour of and Opportunisti- (O-) with that of Round Robin, PFT, and MaxC/I. Video, voie and web traffi are transmitted through an HSDPA air interfae. This is ahieved by means of a ustom built simulator, whih shows how the Earliest Deadline First () sheduler fills an essential role in providing QoS for delay sensitive traffi. II. TRAFFIC GENERATION MODEL The simulation model assumes that multiple soures are produing voie, video, and web traffi that will arrive at the base station, enter the queueing system and is then transmitted via a hannel to the mobile stations. Note that eah physial mobile unit an simultaneously generate voie, video and web traffi, in other words, it an onsist of several traffi soures. The same traffi generating models for video and web traffi will be used as [], while the video generating model will be extended to give a realisti voie generating model. The sheduling intervals are defined in terms of the Transmission This work was partially sponsored by Alatel-Luent and Telkom SA Ltd as part of the Centre of Exellene Programme. Time Interval (TTI). In other words, a new paket is sheduled every TTI. In an HSDPA system, the TTI is fixed at 2ms. A. Video traffi soure A video soure an be modeled as M independent ON- OFF Markov mini-soures. As was the ase in [], this model assumes that M = 0. In the ON state, a mini-soure produes a onstant rate of V bps, while in the OFF state, a mini-soure produes no traffi. Eah mini-soure spends a mean time of p TTIs in the ON state and q TTIs in the OFF state, where their respetive random variables P and Q are geometrially distributed. Parameters p, q and V are obtained as follows: ) p = ( + µ2 a TTI Mσ 2 ) q = ( + Mσ2 a TTI µ 2 V = µ M + σ2 µ [TTIs] () [TTIs] (2) [bps] (3) A video soure has a mean bit rate of µbps, a standard deviation of σbps, and exponent of the auto-ovariane with oeffiient as. As was the ase in [], µ=28kbps, σ=8kbps and a=3.9s. Video pakets are hosen to have a fixed paket length of 500 Bytes = 2kb. Video traffi is onsidered to have a deadline of d=00ms, after whih it will be dropped. The result is that the mean ON-time is p = TTI = 6,820,53 µs (4) Similarly, the mean OFF-time is q = 33.2 TTI = 266,420 µs (5) And finally, the transmission rate of eah mini-soure is B. Voie traffi soure V = 3,300 bps (6) For the voie traffi generation, exatly the same model is used as for video traffi generation, exept that M=. In other words, every soure no longer onsists of 0 mini-soures. For voie traffi, µ=64,000bps and σ=4,000bps. Voie pakets have a fixed paket length of 80 Bytes = 640b. Voie traffi has a deadline of d=50ms, after whih it will be dropped.

2 One again, the mean ON-time is alulated to be p = 32,948.7 TTI = 65,897,436 µs (7) Similarly, the mean OFF-time is given by q = 28.7 TTI = 257,42 µs (8) And finally, the soure transmission rate is C. Web traffi soure V = 64,250 bps (9) Although it is fairly diffiult to find a statistial distribution that aurately desribes web traffi generation patterns, reating simulated web-traffi is fairly straight forward. We use the same model as in [], where an end-user an be seen to osillate between two states while browsing. He is either requesting a new webpage or reading. The reading state has a geometri distribution with an average duration of T OFF =000TTIs = 2s. During this time no traffi is generated. Web traffi has a virtual deadline of d=500ms. It is virtual beause the paket is still served and not dropped when the delay exeeds d and reates a deadline violation. When in the request state, the addressed data soure will produe a geometrially distributed number of pakets, with a mean of P =300 pakets. The paket inter-arrival time too is geometrially distributed, with a mean of T =200TTIs = 0.4s. The paket length L an be found by finding the floor of a trunated Pareto pdf, in other words, L = x, where: p(x) = ζ l min ζ x ζ+ [u(x l min ) u(x l max )] ( ) ζ lmin + δ(x l max ) [bits] (0) l max Here ζ =. is a onstant, u( ) is the unitary step funtion, δ( ) is the Dira Delta funtion, while l min = 80 Bytes = 640b and l max = 500 Bytes = 2kb are respetively the minimum and maximum message lengths. The values of l min and l max are different to those proposed in [] and were hosen to lie within the bounds of the paket lengths of voie and video pakets. The problem with generating web traffi is that the IMSL library that was used to perform statistial tasks does not inlude a funtion for generating Pareto distributed random numbers. To solve this problem, it was noted that the Pareto density funtion is given by p(x) = aba x a+ [bits], () while the Pareto umulative distribution is given by F(x) = ( lmin x ) ζ [bits]. (2) Pareto distributed random numbers an be obtained, by generating random numbers for a uniform distribution with limits [0,]. For every uniformly distributed random u generated in this fashion, F(x) = u. One may then solve for x, as follows: x = l min ( u) /ζ [bits]. (3) The result is that one is able to translate every uniformly distributed random number u into a Pareto distributed random number x. III. QUEUEING MODEL To be able to ompare all shedulers, idential onditions must be reated in eah ase. The only differene may be the way the data is sheduled, in other words, the order in whih it is transmitted. As desribed earlier, traffi that has violated its delay deadline is either physially dropped or virtually dropped, depending on whether it is real-time traffi (voie and video) or best-effort traffi (web). Beause the real-time traffi an be dropped, its queue-length is indiretly limited to the produt of the average servie rate and the delay deadline. Under high enough load onditions, the best-effort queue ould, on the other hand, grow to a signifiant size, whih in turn drastially affets its delay behaviour. Limiting the queue size in any way would also indiretly ap the possible maximum delay. The result is that infinitely long buffers were hosen for all queues. IV. CHANNEL MODEL HSDPA uses Diret-Sequene Code Division Multiple Aess (DS-CDMA). The bit energy-to-interferene spetral density E b that the User Equipment (UE) measures is fed bak to Node-B using 5 bits, known as the Channel Quality Indiator (CQI) value. The CQI value is used to deide on the highest transmission rate that an be hosen. The E b value that the UE measures, an be modeled as follows []: E b = SIR PG, (4) where SIR is the signal-to-interferene power ratio, PG is the DSSS Proessing Gain, whih is defined by the ratio W R b, where W is the spreading bandwidth and R b is the bitrate that depends on the used modulation and oding sheme (MCS). The SIR value an be estimated as SIR = P RX I Inter + I Extra + N, (5) where P RX is the useful reeived power, I Inter is the intraell interferene, I Extra is the extra-ell interferene and N is the noise power. Note that P RX, I Inter, I Extra and N an be haraterized as follows []: G PInter P RX = P TX G PRX, I Inter = αg PInter P D, (6) I Extra = εg IExtra P D, N = N 0 W (7) where P TX is the transmitted power to the user, G PRX, and G IExtra are the user hannel gains of the,

3 transmission, the intra-ell interferene, and the extra-ell interferene, respetively. P D is the Node B available power, N 0 is the noise spetral density, α is the intra-ell nonorthogonality oeffiient and ε is the extra-ell interferene power-to-the total reeived power ratio. Note that all powers must be onverted to Watts and annot be left as dbs. P TX an be found quite simply, by assuming that all of Node B s power will be used during transmissions. Node B s power is therefore divided amongst eah of the transmission odes. If one assumes that 5 odes are used for transmission, then P TX beomes: Hene P TX = P D 5 E b = (αg PInter [W]. (8) P TX G PRX W/R b + εg PExtra )P D + N 0 W (9) per CDMA ode. From [2], one an derive that the hannel gain for the transmission, the intra-ell interferene, and the extra-ell interferene are given by: G (db) = L(t) (db) + s(t) (db), [dbw] (20) where s(t) is the shadowing omponent, whih will be disussed later. L(t) is the path-loss omponent of a UHF signal over flat terrain (in a semi-urban environment), whih is given by [2]: L(t) (db) = K + γ log 0 r dbw. (2) Here K is the signal strength measured at km from the basestation, with r in km, while γ is the rate at whih L(t) (db) hanges as r hanges. Both K and γ may vary as distane r between a mobile and the base-station hanges. The shadowing term s(t) (in db) is usually modeled as a zero-mean stationary Gaussian proess, whih when expressed in Watts has a log-normal pdf: sine µ = 0. f s(t) = σ 2πt exp [ (ln t µ) 2 /(2σ 2 ) ] (22) [ ] = σ 2πt exp ln2 t 2σ 2, (23) V. HSDPA TRANSMISSION RATES When a paket is seleted for transmission to the j-th user, the bit-error-rate (BER) plays an important role. The appliations using the various lasses of servie have different sensitivities to the BER. The BER thresholds that were used in this paper are for video and voie traffi, and for web traffi, as listed in Table II in Setion VII. These values were hosen to orrespond to a paket-error-rate (PER) proposed in [] of 0 for video traffi and a mean PER of 0 2 for web traffi, whih varies as the paket length of the web pakets vary. The BER of the voie traffi is kept the same as video, TABLE I HSDPA MCS MODES MCS mode m Data rate (5 odes) R b. QPSK, rate 4.8Mbps 2. QPSK, rate 2 3.6Mbps 3. QPSK, rate Mbps 4. 6QAM, rate 2 7.2Mbps 5. 6QAM, rate Mbps but orresponds to a muh lower PER, as an be alulated using the following expression: PER = ( BER) L, (24) where L in this ontext is the paket length, measured in bits. The BER an be ontrolled diretly by varying the transmission power. In the 3GPP s UMTS standard, power ontrol therefore plays an important role. One of the improvements that HSDPA offers, is to trak a BER by varying the modulation and oding sheme (MCS) as the hannel onditions hange. This enables the base station to keep its transmission power onstant, simplifying the design. In order to obtain high transmission rates, 5 odes were used with the MCS modes listed in Table I. As the hannel quality varies, a different MCS will have to be used, whih effetively varies the transmission rate. To determine whih of these MCS modes should be used at any time, the indiret relationship between BER and transmission rate needs to be onsidered. Note that the BER is related to the bit-energy-to-interferene spetral density E b value, whih is ommuniated from the User Equipment (UE) to the transmitting Node-B using the Channel Quality Indiator (CQI) value [3]. This 5-bit CQI value, ranging from 0 to 30, is alulated as follows: 0 if E b 6, CQI = Eb / if 4 < E b if 6 < E b 4,. (25) When the CQI value is reeived by Node-B, it an be E onverted bak to an estimate of b. In the previous setion, the following expression was developed that explains the relationship between the transmission rate R b and the resulting E b that an be expeted: E b = (αg PInter P TX G PRX W/R b + εg PExtra )P D + N 0 W (26) By hoosing the appropriate MCS value, the transmission rate an be inreased while the resulting E b still meets the following ondition: ( ) E b Eb >. (27) Threshold ( Eb is a threshold that is diretly related to the )Threshold The maximum PER threshold that eah lass of traffi an endure. Referene [4] ontains the required SNR-to-PER urves in a flat fading Rayleigh model for QPSK rate /3, rate /2, rate 3/4, 6QAM rate /2, and rate 3/4.

4 VI. SCHEDULERS TO COMPARE The behaviour of the following shedulers is explored in this paper: A. Round Robin [3] The Round Robin () sheduler is one of the simplest. At the end of eah TTI, it moves onto the next queue and tries to fill up the sheduling interval with as muh data as possible from this queue. If a queue is empty, the sheduler simply moves to the next queue. No attempt at optimisation is made. If all queues are full, then over a prolonged period of time, eah one would be given an equal number of TTIs for transmission. B. Proportional Fair Throughput [3] The aim of the Proportional Fair Throughput (PF-T) sheduler is to maximise throughput, but in a fair manner. The sheduling rule is given by: Next paket = max i r i r i, (28) where r i is the instantaneous transmission rate that HSDPA ould assign to lass i traffi, if hosen by the sheduler. r i is the average transmission rate that was assigned to lass i traffi. It an be found as follows: { ( α) r i (k) + αr i (k) if i is served in slot k, r i (k + ) = ( α) r i (k) otherwise, (29) where 0 < α < is a weighting fator, whose value was hosen to be The numerator in the sheduling rule ensures that the sheduler will take advantage of temporary throughput improvements. On the other hand, the denominator ensures that over a long-term period, the sheduler will attempt to assign equal resoures to all lasses. C. Maximum Carrier to Interferene ratio [3] The Maximum Carrier to Interferene ratio () sheduler selets pakets based purely on the best hannel onditions. The relative instantaneous hannel quality indiator η is defined by η i = ζ ī ζ, (30) where ζ i = E b is the SNR of the hannel that the head-ofqueue paket of the lass i queue will be transmitted through, while ζ = n n j= ζ j is the average SNR of all hannels. Here n is the total number of traffi lasses. The sheduling rule is simply given by: Next paket = max η i. (3) i D. Earliest Deadline First The Earliest Deadline First () sheduler attempts to meet the required deadlines of traffi lasses. Distributing bandwidth and ahieving fair throughput among the traffi lasses is not a suffiient riterion to be able to sustain realtime traffi on a paketised network. Strit deadlines need to be adhered to. The sheduling rule for is given by: Next paket = min i (d i D i ), (32) where D i is the queueing delay that the head-of-queue paket of the lass i queue has experiened, while d i is the delay deadline of the paket. E. O- Sheduler Finally, Opportunisti (O-) uses the rules of, but prioritises traffi lasses whih will be transmitted through good hannel onditions. Ḡ Next paket = min d i D i (33) i G VII. SIMULATION PARAMETERS Table II lists the parameters for the simulation, whih are similar to those proposed in []. A variable number of traffi soures were modeled, depending on the required load. Eah traffi soure generated the video, voie, and web traffi aording to the model disussed in Setion II. Traffi was generated independently of other traffi soures. For the HSDPA links between Node B and the UEs, 20 mobile hannels were reated. Eah paket that was generated was assigned a flat random number ranging from 0 to 9, whih implied the target hannel that it would be transmitted through. Eah of the 20 hannels was modeled independently of all other hannels. No spaial orrelation among transmission hannels was taken into onsideration. VIII. RESULTS In this paper, the behaviour of five shedulers was simulated. A. Average delay and unfairness Figs. (a), (), and (e) ontain the average queueing delay that video, voie and web traffi respetively experiene. At first glane, seems to perform the least favourable, as of all three lasses it has the highest queueing delay for large portions of the load onditions. One needs to remember though that shedules traffi based on the delay deadlines of the lasses, whih are 00ms for video traffi, 50ms for voie traffi, and 500ms for web traffi. In the ase of voie traffi, traffi does not reah the deadline, but is still muh higher than the other shedulers under high load onditions. The reason for voie delays not reahing the deadline is due to multiple transmissions of voie pakets per TTI. The delay of video and web traffi, on the other hand, approahes the deadline at relatively low loads. The advantage of the sheduler is the lear queueing delay differentiation of traffi lasses. The delay deadlines

5 of the various traffi lasses reate behavioral expetations with ustomers, whih they are prepared to pay for. If no disernable differene is notieable, the requirement for an advaned delay-targeting QoS design beomes questionable. To further illustrate the point, a form of delay unfairness is proposed that is defined as follows: Unfairness = 00% di D i d i, (34) where d i is the delay deadline of lass i and D i is the queueing delay that the front-of-queue-i paket has experiened by the time it is served. In other words, with this expression an attempt is made to measure how lose the delay was to its deadline, by the time the paket ompleted transmission. Figs. (b), (d), and (f) ontain the average unfairness, measured aording to (34). This highlights, more learly, how is able to differentiate the delay of the 3 lasses, based on their respetive deadlines. For all 3 lasses, as the load inreases, approahes the deadline, marked as 0 on the unfairness graphs. In the ase of video traffi, all shedulers are able to approah the delay deadline quite well. For most of the shedulers this is due to their preferene of voie over video traffi. As the video delay approahes the delay deadline, it is dropped. This reates an artifiial eiling around the delay deadline. In the ase of voie and web traffi, all shedulers, exept for, ompletely ignore the delay deadline. IX. CONCLUSION This paper presented an HSDPA simulation environment. Video, voie, and web traffi were realistially produed by a varying number of soures. The number of soures was varied from to 20, thereby ahieving load-onditions varying from 0% to 00%. The model foussed on the downlink transmissions from the base-station to the mobile units, migrating around the ell. It measured the average queueing delay and fairness. The results show how is able to differentiate between traffi lasses. TABLE II SIMULATION PARAMETERS [] PARAMETER VALUE α P D 30 Watt Cell radius R 500 meters N = N 0 W Watt W 5MHz ε (lose to Node B: R 0m) 0 ε (intermediate: 0m < R < 450m) 0.25 ε (ell border: 450m R 500m) 0.5 PATHLOSS ATTENUATION L(t) (in db) Near zone log(r) (R < 300m) Far zone log(r) (R 300m) None-line of sight (probability = 0.2) log(r) (R = distane from Node B) SHADOWING ATTENUATION s(t) (Gaussian variable) (in db) Mean 0 Standard deviation 8 Bit Error Rate Threshold Video Voie Web REFERENCES [] F. D. Angelis, I. Habib, G. Giambene, and S. Giannetti, Sheduling for differentiated traffi types in HSDPA ellular systems, in IEEE Globeom 05, [2] A. Farrokh, F. Blomer, and V. Krishnamurthy, A omparison of opportunisti sheduling algorithms for streaming media in high-speed downlink paket aess (HSDPA), in Pro. of MIPS 2004, Grenoble, Frane, November 6 9, [3] A. Haider and R. Harris, A novel proportional sheduling algorithm for HSDPA in UMTS networks, in The 2nd International Conferene on Wireless Broadband and Ultra Wideband Communiations (I. C. Soiety, ed.), [4] A. Harada, S. Abeta, and M. Sawahashi, Adaptive radio parameter ontrol onsidering QoS for forward link OFCDM wireless aess, in IEEE VTC, pp , Stefan M. Sriba ompleted his BSEng degree in Deember 2000 in the Shool of Eletrial, Eletroni and Computer Engineering at the University of Natal, Durban, South Afria. He ompleted his MSEng degree at the beginning of 2003 and is urrently working on his PhD, researhing sheduling theory for both wired and wireless networks in the Radio Aess Tehnology Researh Centre at the University of KwaZulu-Natal, South Afria. Sine 2006 he is working at Telkom SA Ltd. Professor Fambirai Takawira is the Telkom Professor of Digital Communiations at the Shool of Eletrial, Eletroni and Computer Engineering at the University of KwaZulu-Natal. He ompleted his BSEleEng in Manhester and was awarded a PhD at Cambridge. His researh interests are in the general areas of adaptive signal proessing, digital ommuniations and data networks.

6 Delay (ms) Delay (ms) Delay (ms) (a) Average delay of Video traffi () Average delay of Voie traffi 0 (e) Average delay of Web traffi Unfairness (%) Unfairness (%) Unfairness (%) (b) Unfairness of Video traffi (d) Unfairness of Voie traffi (f) Unfairness of Web traffi Fig.. Average delay and unfairness