Providing Strict Quality of Service in HSDPA for real time services

Transcription

1 Provdng Strct Qualty of Servce n HSDPA for real tme servces Joseph S. Gomes, Mra Yun, Hyeong-Ah Cho, Jae-Hoon Km 2, JungKyo Sohn 3, Hyeong-In Cho 3 Department of Computer Scence, George Washngton Unversty, Washngton, DC 2 Access Network & Moble Termnal R&D Center, SK Telecom, Seoul, Korea 3 Department of Mathematcs, Seoul Natonal Unversty, Seoul, Korea E-mal:{joegomes,mrayun,hcho}@gwu.edu, jayhoon.km@gmal.com, {jgsohn,hcho}@snu.ac.kr Abstract Increasng demand for hgh data-rate real tme servces has led to the use of shared channels n the forward lnk for real-tme servces n 3G wreless networks, such as HSDPA. Strct and dfferentated QoS requrements for real-tme servces, along wth tme varyng channel capacty mpose a great deal of challenge on the MAC-hs scheduler. In ths paper, we present our polcy-based schedulng algorthms to support dfferentated QoS requrements along wth the HSDPA network smulator n OPNET that we developed wth newly ntroduced features, such as lnk adaptaton, HARQ, fast schedulng at NodeB and qualty based admsson control.. Introducton Hgh Speed Downlnk Packet Access (HSDPA) s ntroduced to ncrease WCDMA [] downlnk packet data throughput n the Release 5 of the 3GPP UTRAN specfcatons. HSDPA offers theoretcal peak data rates of above Mbps, whch s acheved by mplementng a fast and complex channel-control mechansm based upon short physcal layer frames (2 ms), Adaptve Modulaton and Codng (AMC), fast Hybrd-ARQ and fast schedulng []. However, generally such hgh data rates and correspondngly hgh throughput cannot be realzed due to shortsghted decsons made by the scheduler. rates (GBR) do not exceed the capacty. Another servce provder may have a dfferent allocaton polcy durng tmes of overload and may choose to mprove channel qualty by ncreasng effectve code rate when there s surplus capacty. In ths paper, we consder such polcy based QoS support [6] [7], present two dfferent algorthms and show that our algorthms very effectvely satsfy the users by guaranteeng such QoS requrements. 2. HSDPA System Overvew Fgure sketches the herarchcal structure of an HSDPA network. In a typcal HSDPA network, data enters the HSDPA/UMTS core network drectly from the Internet or from the PSTN through a Packet Data Swtchng Node (PDSN). In the core-network, data s routed through the GGSN and SGSN, and nto the RNC of the rado access network (RAN). It s then forwarded to the Node-B (base staton) and fnally to the moble staton over the rado lnk. PSTN/ PDSN GGSN SGSN RNC Node B Node B Farness ssues n schedulng have been studed n depth n [2], [3]. The proportonal far (PF) algorthm, most popular packet schedulng algorthm consderng farness, was thoroughly nvestgated n [4], [5], and [6]. Three packet-schedulng algorthms namely Max C/I, Proportonal Farness (PF), and Round Robn (RR) are compared n [7] focusng on the achevable throughput of each user n HSDPA. Reference [8] proposes QoS based schedulng algorthms for real-tme data users over shared wreless lnk. Reference [9] proposes an algorthm that provdes QoS guarantee usng barrer functons and [] consders mult-user transmsson n one slot. A qualty based admsson control algorthm s derved n []. The nteracton between admsson control and fast schedulng have been unexplored under mxed servce requrements n HSDPA networks, although some work has been done n ths area for UMTS networks [2], [3] that use dedcated channels unlke HSDPA. None of the exstng works provdes polcy drven QoS guarantee ether. For example, a servce operator may have a certan polcy on how to prortze QoS classes durng tmes of overload whle another polcy may determne how to dstrbute surplus capacty when the total guaranteed bt Internet Node B HSDPA Cell Fgure : HSDPA Network Archtecture 2. Newly added channels HSDPA Cell Instead of dedcated channels as n release 99, HSDPA uses shared wreless downlnk channels for hgh-speed data. A new downlnk transport channel HS-DSCH (Hgh Speed Downlnk Shared Channel) was ntroduced, that carres data to the selected UE (or UEs) durng each transmsson tme nterval (TTI) of two ms. The transported bts from HS-DSCHs are mapped onto up to 5 physcal downlnk shared channels (HS-PDSCH) each usng a separate orthogonal cdma code wth a spreadng factor of 6. The assocated Hgh Speed Shared Control Channel (HS-SCCH) s used to communcate control nformaton between the UE and Node-B.

2 2.2. Adaptve Modulaton and Codng (AMC) In the uplnk drecton, HS-SCCH channels are used by the UEs to notfy the Node-B of the Channel Qualty Indcator (CQI) and a postve or negatve acknowledgement pertanng to the receved frame. CQI ndcates the nstantaneous channel qualty experenced by the user, so that the Node-B can adjust ts transmsson parameters (modulaton type, codng rate, number of codes) to cope wth varatons n channel condtons. The CQI reported by the UE corresponds to transmsson parameters (modulaton type, codng rate, number of codes) that would result n the mum data rate possble whle provdng an acceptable block error rate (BLER) for the current lnk condtons. The reported CQI value s then used by Node-B to determne the approprate parameters or modulaton and codng set (MCS) for the next packet transmsson to the UE. The actual parameters used are notfed to the UE usng the downlnk HS- SCCH channel. Several AMC schemes are proposed ncludng QPSK and 6QAM wth codng rates of /3, /2 and 3/4. In order to reduce the latency nvolved n lnk adaptaton the schedulng functonalty has been moved from RNC to Node-B. ther GBR. The RNC also receves an E c /N measurement of the CPICH channel from the new user seekng admsson. Based on all these reports, the RNC can estmate whether the user can be granted access wthout deteroratng the servces to the exstng users. (a) Exchangng parameters and (b) Power budget Measurement reports Fgure 2: Admsson Control Parameters 3. OPNET HSDPA Smulator 2.3. UE Categores The UEs are dvded nto 2 dfferent categores based on ther capabltes. For example, UE category k for k = or 2 can only support data rates up to.2 Mbps usng 5 smultaneous physcal channels (codes) and has mnmum nter TTI nterval mn_tti(k) of 3. If user u from category k s scheduled to transmt n TTI t, the earlest TTI n whch u can be scheduled next s t+mn_tti(k). Theoretcally, category 7 can support up to 7.2 Mbps usng codes (under perfect lnk condtons) and has mnmum nter TTI nterval of Hybrd Automatc Repeat Request Hybrd automatc repeat request (HARQ) s a technque where UE stores prevous transmssons that are n error n soft memory to be combned wth future re-transmssons for decodng. For each packet, the UE sends HARQ feedback (ACK or NACK) to nform the Node-B whether a retransmsson s requred or not. HARQ uses one of two dfferent schemes, wth dentcal retransmssons, often referred n the lterature as Chase combnng, or wth non-dentcal retransmssons, otherwse called ncremental redundancy Admsson Control and Schedulng Admsson control s done at the RNC as n release 99, whle schedulng of admtted users s done by MAC-hs scheduler at the Node-B. Resource allocaton, admsson control, and schedulng are collaboratve efforts between the RNC and the Node-B as shown n Fgure 2. The Node-B receves guaranteed bt rate (GBR) and schedulng prorty ndcator (SPI) values from the RNC, whch t can use to make schedulng decsons. On the other hand, RNC allocates channelzaton codes and power for HSDSCH and HS-SCCH transmssons based on load and performance measurements provded by Node-B. These measurements nclude but are not lmted to the total carrer power, non-hsdpa power and the HS-DSCH requred power. HS- DSCH requred power s reported per SPI and s an estmate of the total power needed to serve all admtted user wth that SPI at 2 We developed an HSDPA model by extendng the UMTS model n OPNET. Our system extensvely smulates all the key enttes of the network such as UEs, Node-B, RNC, SGSN etc. wth most of the functonaltes at all the protocol layers. In the followng secton, we brefly descrbe the relevant portons of our smulator mplementaton. A qualty based access control algorthm s mplemented at the RNC, whch wll be descrbed n secton 3. Several packet-schedulng algorthms are mplemented n NodeB ncludng our LPS and SPS algorthms, whch wll dscussed n Secton AMC and HARQ An UE measures the SINR for each receved packet and reports the correspondng CQI back to the NodeB. For ths, actual value nterface (AVI) [4] s used to map a gven modulaton and codng set (MCS) to a receved per-packet average SINR for obtanng a specfc BLER (we use.). For each UE category, we use a separate AVI table that maps an MCS to a correspondng threshold SINR. In Table we show sample MCS to SINR mappngs for ue categores 7 and 8. For example, a CQI of 6 for a UE of category 7 means that the NodeB can transmt 3565 bts to ths UE n the next TTI usng 6-QAM modulaton and an effectve code rate of.37. A UE reports the mum possble CQI for ts current channel condton once every TTI. Error decson for a receved packet s also made based on the threshold SINR for the MCS assocated to the transmtted packet. If the receved SINR s less than the threshold SINR the packet s n error, otherwse a unformly dstrbuted random number y [; ] s generated. If y. the packet s successfully receved. We also model a delay of 2 ms for the Node-B to get the assocated ACK/NACK response back, from the tme of transmsson. We assume Chase combnng for the H-ARQ process and use the followng model [5]: n Es n Es = ε N k = N C, n k

3 , where (E s /N ) C,n represents the combned E s /N after n transmssons and (E s /N ) k corresponds to the k-th transmsson. ε s the Chase combnng effcency, whch s set to.93. In Chase Combnng, an dentcal verson of the orgnal frame has to be retransmtted. Therefore, when mult-user physcal layer frames (code multplexng) are used, even the users that receved ther packet correctly durng the orgnal transmsson wll receve the retransmsson. R k M k = Bk ( m) Ak ( m) (2) N T tt m= Here B k (m) s the transport block sze (TBS) of the m-th transmsson to the user. Notce f the user does not have enough CQI value Transport Block Sze Number of HS-PDSCH codes Modulaton Data rate (Mbps) ECR (per code) SINR (db) Out of Range 37 QPSK QPSK QPSK QAM QAM Table : MCS to SINR mappng table for ue categores 7, 8 Fgure 2: HSDPA RNC process model 3.. Prorty Based Admsson Control at RNC In HSDPA, the schedulng has been moved from RNC to the NodeB. However, the task of admsson control remans n the RNC. The RNC process model n Opnet s depcted n Fgure 2. When an HSDPA packet destned for a partcular UE arrves at RNC n the FROM_CN state, t s forwarded to the servng NodeB. At the begnnng of each sesson, f the necessary resources are not already allocated to the UE, the RNC performs admsson control algorthms n the ADM_CNTL state for both uplnk and downlnk to verfy f enough resources are avalable to serve the UE. The uplnk admsson control algorthm remans the same as n UMTS due to the use of UMTS (DCH channels) n uplnk drecton. For downlnk, we use a prorty based admsson control algorthm that was proposed n [] wth some modfcatons. In ths approach, a new user s admtted only f t can be served wth ts target bt rate wthout degradng the throughput of all the users wth the same or hgher SPI (prorty) from ther target bt rates. We added attrbutes such as sp, downlnk gbr, ue category etc. n the moble staton node model to ndvdually set the schedulng prorty, GBR and category for each UE as shown n Fgure 3. In the prorty-based algorthm, frst the average HS-DSCH power allocated to user k s estmated as, M k ' Pk = Pk ( m) () N m=, where P k(m) s the allocated transmsson power n the m-th TTI where the user was scheduled, and M k s the number of TTIs user k was scheduled out of the N TTIs. When code multplexng s used, we proportonally adjust P k(m) accordng to the multcodes assgned to user k durng that TTI. The average bt rate provded to user k over N TTIs s, Fgure 3: Settng Moble Staton Attrbutes data to send then much of the transport block could be empty and the user wll be experencng a much lower avg. bt rate than the provded bt rate. A k (m) =, f an ack was receved from the user and otherwse. Gven equatons and 2 the requred power to provde all the users, that have SPI value x, wth ther GBR can be approxmated as, GBRk P SPI ( x) = Pk (3) k SPI ( x) Rk, where GBR k s the GBR for user k, and SPI(x) s the set of users wth SPI value x. Now let the target bt rate and prorty of the new user be denoted by GBR new and SPI new respectvely. Let P HSDPA, P new, and P SCCH be the allocated HSDPA transmsson power, the estmated power requred to serve the new user and the estmated power requred for transmttng HS-SCCH respectvely. Then the new HSDPA user s admtted only f + P ( x) + P + P PHSDPA SPI SCCH x SPI new Here P o s a confguraton parameter, whch represents a safety power offset to compensate for potental estmaton errors. In 3

4 [], the requred power to serve the new user at the target bt rate GBR new s estmated as GBRnew Pnew = PTx, DSCH, (5) f ( ρ) where ρ s the average experenced HS-DSCH SINR at the new user as estmated at the RNC from the plot channel measurement report, and f(p) s the average throughput that a user can be served wth when t s scheduled n every TTI wth all the avalable transmsson power. The functon f( ) s assumed to be stored n a table n the RNC. However snce we consder code multplexng we use a slghtly dfferent equaton to compute P new. Suppose the total number of channelzaton codes allocated by the RNC s gven by ψ. Let CQI p be the CQI value that the UE would report when experencng an SINR of p. Let φ p be the number of multcodes correspondng to the transmsson confguraton (MCS) used by the Node-B when sendng to the UE usng CQI p. Then the total transmsson power used to send to the UE can be estmated as P Tx,DSCH (φ ρ /ψ), and thus we have GBR φ new ρ P = new PTx, DSCH (5) f ( ρ) ψ 5. Packet Schedulng at NodeB Fgure 4 shows the process model for an HSDPA NodeB. When a packet destned for a UE arrves from RNC, t s added to the transmsson queue of that user n state HS_ADD_TO_BUFFER. Durng each TTI, n state HS_SCHEDULE, the schedulng canddate set (SCS) s created comprsed of all the users that have non-empty transmsson queues and are ellgble to receve. A user s ellgble, f the number of elapsed TTIs snce ts last recepton exceeds ts mn_tti. Users are chosen from SCS to be scheduled to receve packets based on the schedulng algorthm beng used. connected to t, where each user u s assocated wth a UE category k {,..., 2} and each UE category k s assocated wth mn_tti (k ) {, 2, 3}. GBR denotes the guaranteed bt rate for user u. Node-B s allocated M HS-PDSCH codes by the RNC for downlnk transmsson. CQI (t) CQI = {,,, 3} denotes the CQI value reported by u that can be used for schedulng n TTI t. Gven x CQI, we use blk(x), #CH (x), and DR (x) to represent the block sze (.e., the transport channel bts), the number of requred parallel HSPDSCH codes, and the nstantaneous data rate assocated to x. λ (t) s the assgned CQI for u at TTI t. If u s not scheduled to receve any data at t then λ (t) =. Suppose user u entered the system at TTI s. Then, δ = t - s + s the number of TTIs the user has been n the system. We denote th (t) to be the movng average of the throughput of user at TTI t over the last w o TTIs defned as follows: DR( γ ( t)) th ( t )( ) +, δ w w w th ( t) = t = DR( γ t j s ( )), otherwse t s Schedulng Problem Statement For every TTI t fnd X (t) = (γ (t),,γ n (t)) such that, (a) γ ( t) CQI ( t) for all, n (b) = # CH ( γ ( t)) M, t (c) k = t mn_ TTI ( C ) + γ ( t) = In other words, the goal of the schedulng problem s to select a set of recevng users and to select the nstantaneous transmsson data rates (format) for them such that each user s assgned CQI s less than ther requested CQI, the total number of codes used s less than M, and each selected user s ellgble. Fgure 4: HSDPA NodeB process model 5.. Notatons and Assumptons We make the followng notatons and assumptons to formulate the schedulng problem: a Node-B has n users, u u n, Loose Polcy Schedulng In addton to the above constrants, a servce provder may have other polces that mpose supplementary requrements on the scheduler. These polces could be loose or strct n nature. Here, we defne a loose polcy scheduler to be a scheduler that tres to fnd a balance between the two goals of mzng throughput (usng opportunstc schedulng) and satsfyng QoS constrants (GBR). Notce that ths s dfferent from the goal of PF algorthm, whch dsregards QoS constrant. The PF (proportonal far) scheduler orders the recevers usng the metrc r (t)/th (t), where r (t) s the nstantaneous data and th (t) s the current throughput. It consders all users equally mportant and adds a farness property wth respect to user throughput. We try to ncorporate both QoS constrants and throughput mzaton nto our algorthm by consderng the DR( γ ( t)) GBR metrc μ =. Notce that the frst rato # CH ( γ ( t)) th ( t) contrbutes to the users wth better nstantaneous data rate per code and the second rato emphaszes users who are furthest

5 away from meetng ther guaranteed bt rate. Next, we gve the algorthmc steps n further detal.. Let U actve (t) denote the set of users who may be scheduled for transmsson n TTI t (.e., constrant (c) s satsfed.) For each user u U actve (t), defne Δ (t)= w (GBR - Th (t -)) + Th (t - ). 2. Gven CQI (t), choose smallest γ (t) such that () γ (t) CQI (t) and () DR(γ (t)) Δ (t). 3. Sort the users n U actve (t) n descendng order of μ. 4. Let z denote the number of unused HSPDSCHs. We are done when ether z = or U actve s empty. Otherwse, choose the user from the sorted lst wth the hghest μ and assgn #CH(γ (t)) codes to the user f #CH(γ (t)) < z. Otherwse, assgn z codes and the hghest possble CQI usng z codes to the user. Remove the user from U actve. 5. Repeat Step Strct Polcy Schedulng Now we wll consder a strcter set of polcy requrements. Suppose a servce provder dfferentates ts users usng Schedulng Prorty Indcator (SPI) values, where SPI C = {, 2,, 5}[]. In general, we assume hgher classes (wth hgher SPI) have hgher GBR values. Although the problem can be generalzed for any number of SPI values, for smplcty let us consder only SPI and 2 wth correspondng GBR values GBR s and GBR g. We wll call these two classes Gold and Slver classes respectvely. Let d (t) denote the mum achevable nstantaneous data rate ndcated by the CQI for user u at TTI t, where d (t) {r,r 2,..., r k } such that r = r mn < r 2 <... < r k = r. We assume that a user belongng to a partcular class s on average under good enough channel condton to be able to receve ts GBR. Under ths assumpton, the goal of our polcy s to defne far rules for governng resource allocaton under all crcumstances,.e. to guarantee each user ts GBR and farly dstrbute the surplus capacty when there are enough resources, and to satsfy users from hgher classes before the lower classes when there are not enough resources. The polces are as follows: P. A slver user can be scheduled only f all gold users have been satsfed. P2. If there are multple gold users wth throughput less than GBR g, the gold user wth the hghest value of d (t) s scheduled. P3. A slver user s wth th s (t) less than GBR s has hgher prorty than any gold or slver user that meets ts GBR. P4. If there are multple slver users wth throughput less than GBRs whle all gold users have been satsfed, a slver user wth the hghest value of d (t) s scheduled. P5. When all users have met ther GBR, surplus capacty must be proportonally dstrbuted among the gold and slver users accordng to ther GBRs Margnal Utlty Functon t accordng to ts class c, current data rate d (t) and current throughput th (t). Eventually the utlty values wll be used to determne whch user(s) wll be scheduled at t. Let M c (th (t)) = P c (th (t)) d (t) where P c denotes the prelmnary utlty functon whch we wll defne shortly. Then for any gold user, g, and slver user, s, P Gold and P Slver has to follow the followng condtons: C. P Gold (th g (t)) r mn > P Slver (th s (t)) r for th g (t) < GBR g and th s (t) r C2. P Slver (th s (t)) r mn > P Gold (th g (t)) r for th s (t) < GBR s and GBR g th g (t) r GBR C3. P Slver (th s (t)) = P Gold ths ( t), for GBR s th s (t) th ( t) r and GBR g th g (t) r. C4. For any two slver users, s and s, P Slver (th s (t)) r mn > P Slver (th s (t)) r for th s (t) < GBRs and GBR s ths (t) r C5. For any two gold users g and g, P Gold (th g (t)) r mn > P Gold (th g (t)) r for th g (t) < GBR g and GBR g thg (t) r. Note that Condton C and C5 correspond to polcy P, C2 and C4 correspond to P3, and Condton C3 corresponds to P5. We defne functons P Gold and P Slver satsfyng the condtons above as follows. Let β = r - GBR g, α = r /r mn and x be the throughput of the user. P P Slver Gold αβ +, x < GBR ( x) = GBRg r x, GBRs x r GBRs 2 α β + α +, x < GBRg ( x) = r x, GBRg x r Functons P Gold and P Slver are shown n Fg.. Notce that mum P Slver value for a slver user wth throughput th s (t) GBR s and mum P Gold value a gold user g wth throughput th g (t) GBR g s β. So the mum margnal utlty (M c ) for these users s βr. The mnmum margnal utlty for a slver user s wth th s (t) < GBR s s (αβ + ) rmn > βr. Ths satsfes condtons C2 and C4. Smlarly f th s (t) < GBR s, M Slver (th s (t)) (αβ + ) r. For a gold user wth th g (t) < GBRg, M Gold (th g (t)) (α 2 β + α + ) r mn (αβ + ) r. Ths satsfes condtons C and C5. It can also be notced from the fgure that when all users meet ther GBR, condton C3 s also satsfed. Also snce two gold users wth throughtput less than GBR g has the same P value, the one wth the hgher data rate wll have a hgher margnal utlty, whch conforms to P2. For smlar reasons P4 s also satsfed. s Let N be the set of users n the system who can receve at TTI t. We wll defne margnal utlty functons M c (th (t)) for each class c {Gold, Slver}, whose purpose s to assgn a utlty value, followng the polcy rules descrbed earler, to each user u at TTI 5 Suppose there are more than two classes,.e. C > 2. Let GBR denote the GBR assocated to the hghest class or SPI. Then P c can be generalzed for any SPI value c C as

6 c c α β + α + K+ α, P ( ) = GBR c x r x, GBRc Strct Polcy Schedulng Algorthm x < GBR GBR c x r Suppose U s the lst of users that are ellgble to receve n TTI t. A user s ellgble f there are packets to be sent to that user and mnmum nter TTI nterval for that user has elapsed snce ts last recepton. Let M be the number of codes allocated for the HS-DSCH channels and codes left be the number of codes left to be assgned. Durng each TTI t, the followng steps are used to produce the lst of users that are scheduled to receve data.. Let be the user such that = arg j {M cj (th j (t))} where user j belongs to class c j. Let cq be ts requested CQI, codes(cq ) be the number of codes requred. a. If (codes (cq ) < codes_left) Send to user u usng data rate DR (cq ) Else Send to user u at the mum rate possble Usng codes_left number of codes. b. Update codes left and remove from U. 2. If (codes_left > and U Ø) goto Step. c Shadow fadng std. Ste-to-ste dstance User speed Throughtput wndow DB 2km -2 km/h TTI Table 2: Summary of Man Smulaton Parameters 6.. Evaluatng Loose Polcy Schedulng Algorthm We compared our LPS algorthm wth Max C/I and PF, usng the followng scenaro. There are ten gold and ten slver users all wth UE category 7. On average, the gold users are at a worse (by around 5 db) rado condton than the slver users. Ten codes are allocated at NodeB for downlnk transmsson. Each gold user has a guaranteed bt rate of 384 kbps, whereas for the slver users t s 28 kbps. Packets are beng generated destned to each gold and slver user at the same rate as ther GBR. Average Throughput (kbps) Max C/I (Slver User) Max C/I (Gold User) LPS (Slver User) LPS (Gold User) Tme (Sec) Fgure 6: Comparng Loose Polcy Scheduler and Proportonal Farness 45 PF (Slver User) PF (Gold User) LPS (Slver User) LPS (Gold User) 6. Performance Evaluaton Fgure 5: P Gold and P Slver In order to evaluate our proposed schedulers we smulated several scenaros. The GBR for each class, the traffc load, and capacty of the network (HS-PDSCH codes) are vared based on the scenaros. The man parameter settngs of our smulatons are shown n Table I. Parameter Settng Total HS-PDSCH power 8 W Number of HS-PDSCH codes 5, H-ARQ Chase Combnng Number of H-ARQ channels 6 Max number L transmssons 4 CQI reportng nterval On every TTI HSDPA termnal category, 7 Path loss model Vehcular Outdoor 6 Average Throughput (kbps) Tme (Sec) Fgure 7: Comparng Loose Polcy Scheduler and Max C/I Fgures 2 and 3 show the average throughputs of a typcal gold and slver user. Notce n Fgure 2 the PF scheduler provdes the slver users wth a hgher throughput (around 28 kbps) than the gold users (75 kbps), snce they have better channel condton. At every TTI, Max C/I (Fg. 3) always choose the UE wth best

7 channel condton. Consequently, the slver users receve at 28 kbps, whereas the gold users are much starved. On the other hand, LPS meets both gold and slver users throughput requrements. Moreover, snce PF and Max C/I choose one user for each TTI, the capacty cannot be fully utlzed f the scheduled user s queue does not have enough data to match the chosen CQI. Snce the slver users, that have lesser data destned for them, are beng chosen most of the tme (always by Max C/I), the cell capacty s beng under utlzed. On the other hand, as LPS chooses the mnmum CQI that supports the GBR, t can send to multple users at the same tme, when M s large enough. Ths, together wth the hgh GBR of the gold users, ncreases capacty utlzaton. Notce n Fg. 4 that LPS has an average downlnk throughput of 5 Mbps compared to Max C/I s.28 Mbps and PF scheduler s 2 Mbps. receve smlar throughput (5 kbps) as they are experencng smlar rado condtons (SINR). Throughput (kbps) Slver User Exstng Gold User New Gold User Total Throughput (Mbps) Max C/I PF LPS Tme (secs) Fgure 8: Comparng total receved throughput at the moble statons In a smlar scenaro, we moved one of the gold users further away from the NodeB than the rest of the users to worsen ts rado condtons (2 db). Consequently, LCS reduced ts average throughput to around 2 kbps, whle stll mantanng the GBR for all the other users. Ths ndcates that LCS can relax the QoS constrants n order to ncrease throughput. 6.. Evaluatng Strct Polcy Schedulng Algorthm ) Scenaro : In ths scenaro, GBR g and GBR s are set to 25 kbps and kbps respectvely. The traffc load destned to the Gold and Slver users are 3 and 5 kbps. 5 codes were allocated for the HS-PDSCH channels. A new gold user s added to the network after 5 seconds. For UE category the downlnk capacty usng 5 codes s around 3.6 Mbps under good channel condtons. From Fg. 5 we can see that before addng the new user, all users are recevng at ther GBR and the total downlnk throughtput n the cell reaches cell capacty. After the new gold user came n ts throughput quckly went upto 25 kbps; however each slver user s throughput went down by about 25 kbps to make room for the new gold user. However, Fg. 6 llustrates the dsablty of PF Scheduler n mantanng any QoS requrement. In ths case, both gold and slver users 7 Throughput (kbps) Throughput (kbps) Tme (Sec) Fgure 9: Change n throughput when extra user added Gold User Slver User Tme (Sec) Fgure : Gold and Slver throughput wth PF Gold User Slver User Tme (Sec) Fgure : SPS throughput when traffc load exceeds GBR 2) Scenaro 2: For ths case, GBR g and GBR s were set to 28 and 64 kbps respectvely. The traffc load destned to the Gold and Slver users were 6 and 8 kbps. Ten codes were allocated

8 for the HS-PDSCH channels. Notce that both users are recevng at a hgher rate than ther GBR at all tmes. The addtonal throughput s beng dstrbuted proportonally after satsfyng all users guaranteed bt rates. 3) Scenaro 3: In ths scenaro, we ncrease GBR s for gold and slver users to 256 kbps and 28 kbps respectvely and set the number of codes allocated to fve to create a shortness of resources. The traffc load destned to the Gold and Slver users were 3 and 5 kbps. In Fgure 8, notce that the gold user s recevng at ts requred GBR at all tmes. However due to capacty constrants the slver user s recevng at a lower rate (93 kbps) than ts GBR snce the total combned throughput reaches the cell capacty. Throughput (kbps) Slver User Gold User Tme (Sec) Fgure 2: User throughput n SPS when total requested GBR exceed capacty 7. Concluson We developed and nvestgated two schedulng algorthms for HSDPA systems based on loose and strct enforcement of QoS aware polcy constrants. We compared our algorthms aganst well-known schedulers usng the HSDPA smulator that we developed n OPNETand showed that both algorthms perform sgnfcantly better n terms of supportng QoS constrants. They also sgnfcantly ncrease system throughput by schedulng multple users n each slot. Our loose polcy scheduler can also relax the QoS enforcement under bad rado condtons to ncrease throughput. References [] H. Holma and A. Toskala, Eds., HSDPA/HSUPA for UMTS. Wley, Aprl 26. [2] T. Koldng, F. Frederksen, and P. Mogensen, Performance aspects of wcdma systems wth hgh speed downlnk packet access (hsdpa), n Vehcular Technology Conference, 22, pp [3] X. Lu, E. Chong, and N. Shroff, A framework for opportunstc schedulng n wreless networks, Computer Networks Journal, vol. 4, no. 4, pp , 23. [4] F. Kelly, Chargng and rate control for elastc traffc, European Transactons on Telecommuncatons, vol. 8, pp , January 997. [5] A. Jalal, R. Padovan, and R. Pankaj, Data throughput of cdma-hdr a hgh effcency-hgh data rate personal communcaton wreless system, n Vehcular Technology Conference Proceedngs, 2. VTC 2-Sprng Tokyo. 2 IEEE 5st, vol. 3, pp [6] T. Koldng, Lnk and system performance aspects of proportonal far schedulng n wcdma/hsdpa, n Vehcular Technology Conference, 23. VTC 23-Fall. 23 IEEE 58th, vol. 3, pp [7] Y. Ofuj, A. Mormoto, S. Abeta, and M. Sawahash, Comparson of packet schedulng algorthms focusng on user throughput n hgh speed downlnk packet access, The 3th IEEE Internatonal Symposum on Personal, Indoor and Moble Rado Communcatons, vol. 3, pp , 22. [8] M. Andrews, K. Kumaran, K. Ramanan, A. Stolyar, P. Whtng, and R. Vjayakumar, Provdng qualty of servce over a shared wreless lnk, Communcatons Magazne, IEEE, 25., vol. 39, no. 2, pp. 5 54, Feb. 2. [9] P. Hosen, Qos control for wcdma hgh speed packet data, n 4th Internatonal Workshop on Moble and Wreless Communcatons Network, 22, pp [] F. Berggren and R. Jantt, Multuser schedulng over raylegh fadng channels, n Proc. IEEE Globecom, December 23. [] K. Pedersen, Qualty based hsdpa access algorthms, n Vehcular Technology Conference, 25, pp [2] S. Elayoub, T. Chahed, and G. Hbuterne, Resource management n umts: Admsson control and packet schedulng, n ICON 23, October 23. [3] S. Pal, M. Chatterjee, and S. K. Das, Call admsson control and schedulng polces for umts traffc for qos provsonng, n Hgh Speed Networks and Multmeda Communcatons 7th IEEE Internatonal Conference, HSNMC 24, 24. [4] S. Hmlnen, P. Slanna, M. Hartman, A. Läppetelnen, H. Holma, and O. Salonaho, A novel nterface between lnk and system level smulatons, n ACTS Moble Telecommun, Oct. 997, pp [5] F. Frederksen and T. Koldng, Performance and modelng of wcdma/hsdpa transmsson/h-arq schemes, n Vehcular Technology Conference, 22, pp [6] J. S. Gomes, M. Yun, H.-A. Cho, J.-H. Km, J. Sohn, and H. I. Cho, Schedulng algorthms for polcy drven qos support n hsdpa networks, n IEEE 65 th Vehcular Technology Conference, 27. VTC 27-Sprng. [7] J. S. Gomes, H.-A. Cho, J.-H. Km, J. Sohn, and H. I. Cho, Integratng admsson control and packet schedulng for qualty controlled streamng servces n hsdpa networks, Accepted for publcaton n IEEE Broadnets, 27. 8