Improving Load Balancing mechanisms in Wireless Packet Networks 1

Transcription

1 Improvng Load Balancng mechansms n Wreless Packet etworks 1 Guseppe Banch, Ilena Tnnrello Dept. Electrcal Engneerng, Unversty of Palermo Vale Delle Scenze, Parco D'Orleans, Palermo, Italy Abstract - Ths paper provdes a comparatve performance evaluaton of varous load balancng schemes n cellular packet networks. Wth respect to crcut swtched networks, wreless packet technology adds the further ssue of Qualty of Servce of accepted connectons. In fact, wth packet technology, transmsson error performance does not unquely depend on the perceved channel qualty, but t can be mproved by adoptng a schedulng mechansm enforcng fast retransmsson of corrupted packets. The result s that throughput can be traded off wth QoS experenced by an admtted flow. Ths paper proposes new packet-level load balancng mechansms. In addton to the number of calls admtted n a cell, our schemes use supplementary packet level nformaton, expressed n terms of effectve resource consumpton of each ndvdual call when retransmsson mechansms are employed. Smulaton results prove the superorty of our proposed schemes wth respect to tradtonal load balancng schemes. I. ITRODUCTIO Tradtonal frequency and tme dvson cellular systems are based on fxed channel assgnment. In such systems, a fxed channel resource, e.g. a TDMA slot per frame, s allocated to an admtted call for ts whole duraton tme. It results that the QoS offered to the users s summarzed n two dstnct and ndependent performance fgures: transmsson qualty, whch depends on the transmsson error probablty encountered as well as on the specfc error correctng codng employed, and outage probablty (.e. the probablty that a call s refused or subject to a forced nterrupton). Load balancng algorthms [1,2,3,4] have been consdered to mprove the outage probablty n wreless cellular networks. These algorthms operate when the coverage areas of dfferent Base Statons (BS) overlap. Whenever a Moble Staton (MS) can attach to more than one BS, the dea s to drect a new call to the BS wth the greatest number of avalable channels. It has been proven that ths operaton mnmzes the probablty that future ncomng calls (ether newly orgnated or handoverng ones) wll be blocked because of lack of resources. Wreless networks based on packet technology and dynamc resource assgnment [5,6,7,8] are ncreasngly popular. In these systems, channel access s dynamcally controlled by a packet (slot) scheduler runnng at the BS, whch grants channel access to the MS wth most urgent need to transmt. Dynamc resource assgnment s partcularly effectve as t provdes the capablty to retransmt, n a very short tme frame, corrupted packets, thus mprovng the packet error rate n the presence of crtcal channel condtons [5,6]. Emergng commercal WLA systems, such as Bluetooth, and the Pont Coordnaton Functon of the IEEE , have ncluded dynamc packet schedulng n ther channel access control operaton. Fast ARQ s also envsoned n thrd generaton cellular systems (UMTS). The thess carred out n ths paper s that, n wreless packet networks, load balancng algorthms should be desgned to take advantage of addtonal "packet level" nformaton, rather than be lmted to "call level" nformaton (number of accepted calls per cell). In fact, n dynamc resource assgnment schemes, the transmsson error probablty,.e. the packet level QoS, s no more ndependent on the cell occupancy status. On the contrary, packet retransmsson capabltes mply that the greater the amount of spare channel resources, the larger s the number of retransmsson that can be attempted, the better the transmsson qualty provded. Scope of ths paper s to propose novel load balancng schemes drven by packet level nformaton, and prove ther superorty wth respect to tradtonal load balancng approaches. More nto detals, we propose and evaluate the performance of two novel metrcs, based on packet level nformaton, desgned to drve the BS selecton algorthm for newly orgnated or handoverng call. The paper s organzed as follows. In Secton II, dynamc resource assgnment systems are brefly descrbed. Secton III proposes and motvates the novel BS selecton metrcs. In secton IV, a performance evaluaton s carred out by means of smulaton. Fnally, conclusve remarks are drawn n secton V. II. DYAMIC RESOURCE ASSIGMET In a mcro cellular envronment, the small propagaton delay allows an effcent command/response communcaton mode. Ths feature has been effectvely used n several recently proposed multple access control (MAC) schemes, whch base ther operaton on dynamc resource assgnment [5,6,7]. In these schemes, the channel s dvded nto fxed sze slots. However, they dffer from fxed TDMA access mechansms as each slot s not unquely reserved to an attached connecton, but s dynamcally assgned by a central scheduler placed on the BS. Dynamc assgnment s mplemented by a command-response protocol, ether on a slot-by-slot [5,6] or a frame-by-frame [7] bass. Dynamc resource assgnment schemes appear effectve when mult rate and/or varable rate traffc sources, wth dfferent delay requrements and prortes, need to share the same medum. However, t has been proven [5,6] that a flexble slot assgnment turns nto an effectve performance enhancement. Ths work was supported n part by MURST /02/$ IEEE 891

2 Slots used for frst tx: 1st cycle Slots used for retransm: 2nd cycle 3rd cycle A) Slots used for frst tx: 1st cycle retx: 2nd cycle?. B) Fg. 1. Graphcal llustraton of the dynamc resource assgnment scheme adopted. Each connecton s dentfed by a dfferent number. Unsuccessful transmssons are marked. The dea s that dle slots can be used to retransmt corrupted packets: the greater the maxmum packet delay and the spare capacty, the greater the probablty of a successful packet transmsson, even over degraded channels. Fgure 1 llustrates the smplfed dynamc resource assgnment scenaro used n ths paper. We have lmted our nvestgaton to homogeneous constant rate traffc sources. Each wreless cell can accommodate up to M connectons, by means of a TDMA frame of sze M. Each attached source s assumed to generate a new packet at the begnnng of a frame. A packet s lost f t s not successfully transmtted by the tme the next packet arrves,.e. by the end of the frame. As graphcally llustrated n fgure 1, the schedulng algorthm s dvded nto «cycles». The BS gves a frst transmsson opportunty to all connectons. Then, a second transmsson grant s gven to the connectons that have faled the frst transmsson, and so on untl the end of the frame. To guarantee farness, f the number of connectons n the last cycle s greater than the number of avalable slots, random selecton s enforced. In essence, retransmsson s a way to trade off throughput, channel qualty, and loss performance. In fact, resources must be reserved n each cell for packet retransmssons. The amount of ths resources depends on the packet loss target, but also on the channel qualty perceved by the users. The graphcal example shown n Fgure 1 clarfes these consderatons. In the case (A), to guarantee an error-free packet transmsson, a greater number of slots s reserved for retransmssons, because of bad channel condtons. Consequently, the throughput of the cell s low. In the case (B), better transmsson condtons allow to ncrease the throughput, snce the bandwdth reserved for corrupted packet recovery can be reduced. III. "PACKET LEVEL" LOAD BALACIG SCHEMES Let's now dscuss the ratonale that drves the necessty of novel load balancng algorthms. Consder the scenaro dected n fgure 2. Two adjacent BSs serve, respectvely, four and three calls n progress. At a gven nstant of tme, a new MS wants to set up a new connecton (or, equvalently, requres handover nto one of the two cells).load balancng enter nto play when the new MS s capable of attach to ether BSs,.e. t measures a comparable channel qualty to both target BSs (n the fgure, ths s enlghtened by havng postoned the MS at the same dstance wth respect to both BSs). Cell A Cell B Fg. 2. The problem of selectng the less loaded BS If the load balancng operaton s drven by call level nformaton, the new MS would attach to cell B because t s the less loaded cell n terms of actve calls n progress. However, we argue that such decson may lead to neffcency: although, after the MS attaches to cell B, both cells result to accommodate the same number of calls, the load at the packet level may be sgnfcantly dfferent. In fact, snce the MS s n cell B are placed at a much larger dstance from the BS wth respect of the case of cell A, they wll lkely suffer of worse channel condtons and consequently of a greater packet error probablty. Thus cell B wll envson a large extra load generated by packet retransmsson and, n turns, a correspondng degradaton of the packet recovery probablty for the attached statons. It may result that a much better admsson strategy would consst n admttng the MS to cell A, deste the fact that ths would create a load unbalancng at the call level. To ncorporate the above descrbed phenomenon nto a practcal load balancng algorthm, t now results necessary to defne quanttatve rules (metrcs) that allow to select the less loaded cell at the packet level (n the above example cell A nstead of B). The dea s to account, n the cell selecton metrc, nformaton related to the packet level load. To ths purpose, n the followng sectons III.B and III.C, we propose two dfferent metrcs to quantfy such nformaton. The frst metrc s based on the computaton of the average number of packet transmssons wthn a cell. The second metrc s more complex, and attempts to drectly estmate the packet loss performance, whch n turns represents an ndrect measure of the packet load. A. Channel Error Model Our metrcs have been ntroduced n the assumpton of uncorrelated errors.e. the result of consecutve packet transmssons for the same MS are ndependent, and the error probablty depends on the MS dstance to the target BS. However, we have verfed that the conclusons drawn from our study are stll vald when error correlaton s consdered. We descrbe packet successes and falures due to fadng and nose wth a propagaton model that takes nto account Raylegh fadng, due to multpath, and η-th power loss law. The power P R, receved at the BS from a MS located at dstance r, s gven by: P R =α 2 A r -η P T (1) 892

3 where α 2 s an exponentally dstrbuted random varable wth unt mean that represents the fadng, Ar -η accounts for the power loss law, η tycally takes values n the range 3,4 and P T s the transmtted power, whch s consdered the same for all statons. Assumng a threshold model for the packet success, the probablty p r (r) that a packet transmsson by a staton at dstance r s ncorrectly receved at the BS s gven by: PR pr ( r) = Pr b (2) W where W s the total nose power at the BS (assumed constant), and b s the capture rato. Gven W, the parameters A and P T, and recallng that α 2 s exponentally dstrbuted, we have: r η br / SR0 p ( r) = 1 e (3) where SR 0 =AP T /W s the average Sgnal To ose Rato at the cell border. We consder that dfferent users experence dfferent multpath and nose effects, so that ther transmsson processes are ndependent. A more realstc channel error model has been consdered n the smulaton program. Error correlaton has been accounted for wth a two state good/bad Markov model. Ths model assumes that the channel fluctuates between two states: the good state n whch no packet error occurs, and the bad state n whch every packet transmsson fals. To account for the effect of the MSs dstance from the BS, we have mposed that the probablty of fndng the channel n the bad state at a random nstant s p r (r), gven n eq. (3). The model s then completed once the average length of a burst of consecutve errored channel slots, Q, s specfed. We have assumed Q to be ndependent of r. B. The Gross Load Metrc The "Gross Load" (GL) metrc stems from the observaton that, assumng unlmted resource avalablty, the number of packet retransmssons necessary to successfully transmt a sngle packet s a geometrcally dstrbuted random varable,.e: j PES ( j) = (1 ) p (5) s the probablty of a successful packet transmsson after j falures (n other words, the probablty that the -th MS requres j extra slots to complete the packet transmsson). In the above equaton, p s the transmsson error probablty suffered by the -th MS. Ths mples that, n average, the -th MS needs g slot to transmt ts packet successfully, where: g = 1 + E[ extra slots] = 1+ (6) 1 We defne g to be the Gross Load offered by MS. Beng the number of attached MSs, the total Gross Load GL offered to a cell s gven by: GL = g = + (7) = 1 = 1 1 GL s gven by the sum of two terms. The frst s the number of attached MS, that s a statc value when no MSs attach/detach. The second term s a dynamc value, dependng not only on the number of MSs, but also on ther physcal poston, snce p depends on the (tme-varyng) MS dstance r through equaton (3). C. The Metrc The "" metrc s motvated by the observaton that the best possble load balancng metrc s to select the target cell as the one whch mnmzes the expected packet loss percentage after the addton of a new MS. In the smplstc case of all MSs sufferng the same packet error probablty p (.e. placed at the same dstance from the BS), a closed form expresson of the packet loss probablty dstrbuton s gven by: l+ M l M Pr{ lost packets= l} = p (1 p) (8) l where M s the number of avalable slots, and s the number of statons. Thus, the average packet loss n a cell of sze M=+k, wth statons connected, s gven by: + k l l+ loss (, k) = l(1 p) p (9) l = 1 l P 1 Although, an explct loss formula for MS havng dfferent packet error probabltes appears non trval, we have verfed that an excellent approxmaton s to compute the (PL) formula (9) usng, nstead of p, p equ = 1 (10) GL The ratonale at the bass of our approxmaton s to substtute the actual MS dstrbuton wth another one, n whch we consder an equal number of equdstant statons that offer the same gross load GL. D. Load balancng algorthm Once a metrc s selected, the load balancng algorthm s trvally specfed by the followng subsequent steps. Whenever a staton wants to setup a new connecton, or requres an handover n an adjacent cell, t frst determnes the set of BSs n ts coverage area. Second, for each BS n ths set, the MS verfes that at least one BS s able to accommodate the ncomng call (otherwse, the call s refused). The defnton of a sutable connecton admsson control (CAC) scheme s a complex problem, and out of the scopes of the present paper. However, to avod cell overload, we have adopted a tradtonal "call count" based CAC,.e. a call s blocked whenever the number of already admtted calls s equal to M, the number of slots avalable wthn a frame. Fnally, the MS selects the BS target as the one that mnmzes the selected metrc, computed ncludng the contrbute of the ncomng MS. IV. PERFORMACE EVALUATIOS In ths secton we detal the applcaton of the above concepts to a multcellular network and nvestgate the performance of our new load balancng metrcs by means of a 893

4 1,1 2,1 3,1 1,2 2,2 3,2 1,3 4,2 2,3 4,2 4,3 4,4 3,3 1,4 4,3 2,4 3,4 1,1 4,4 2,1 3,1 Fg. 3. etwork model used n the smulaton program smulaton model. The performances are evaluated n terms of packet loss percentage and call blockng probablty offered to the users. A. etwork Model Descrpton To avod border effects n the smulaton results, we have developed a torodal network model (a smlar model has been used n [5]). The network s dected n fgure 3, and conssts of 16 hexagonal cells lyng on a torus surface. As shown n the fgure, each cell always has 6 adjacent cells,.e. there s no network border, e.g. cell (1,3) s adjacent to cell (4,3) and (4,4). The BSs are located at the center of the hexagonal cells and operate wth omn-drectonal antennas. Ths mples that the coverage area of each one s crcular and contguous area overlap. We call overlap factor (OF) the rato between the maxmum dstance d max at whch an MS can attach, and the radus R of the crcle n whch the hexagonal cells are nscrbed. Unless otherwse specfed, the value d max = 1.5 R has been adopted. We have consdered a smulaton scenaro n whch new MSs actvate and deactvate dynamcally. In partcular, the arrval process of new calls has been modeled as a Posson process, whle the poston of the new MS has been unformly drawn on the whole torus surface. Exponental call holdng tmes have been adopted (the average call duraton has been set to 5000 frames). Although no moblty model has been consdered n ths paper, the process of handover from one cell to another can ndeed occurs, f, at a gven nstant of tme, a better QoS s offered n an adjacent cell. In fact, snce the cells sgnfcantly overlap, a user placed n the cell perphery can be served by two or more BSs. In prncple, an handover event to a dfferent cell may be trggered even for statons placed very close to a BS, when, occasonally, the adjacent cells are very low loaded and thus provde better QoS performance. However, ths s not recommended, as future calls attachng to the far cell wll very soon trgger an nverse handover. Thus, we have ntroduced a mnmum threshold, called d trg that we set equal to 3 / 2 * R. Statons whose dstance from a BS s lower than d trg cannot perform handover. When the dstance from the BS becomes greater than ths trggerng threshold, the moble begns to contnuously montor the neghbor BSs. If the BS that mnmzes a gven metrc s dfferent from the current one, an handover occurs. 1,E-05 1,E-06 1,E-07 PL L MD GL 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 Offered Load Fg. 4. Packet loss performance versus offered load, for the dfferent metrcs: SR 0 /b=5, M=20 slots. Block Probablty PL L MD GL 0,4 0,5 0,6 0,7 0,8 0,9 1 Offered Load Fg. 5. Call blockng performance: SR 0 /b=5, M=20 slots. B. umercal Results In ths secton, we compare the performance of our proposed packet-level load balancng metrcs, namely GL and PL, wth the performance provded by the followng two tradtonal approaches Mnmum Dstance (MD). the target BS selected s the closest one (no load balancng s mplemented). omnal Load (L): among the BSs that can admt the user, the one that accommodates the lowest number of calls s selected. Fgure 4 shows the average packet loss performance versus the normalzed per cell offered load, for the four dfferent metrcs. The fgure shows that packet level metrcs are always superor. Specfcally, for offered loads leadng to a packet loss target lower than 10-2, our packet level metrcs acheve an mprovement of at least one order of magntude. A comparson between PL and GL shows that PL provdes better performance, especally when lght load condtons are consdered. The reason s that the PL metrc drectly optmzes the expected packet loss, whle the GL metrc s only ndrectly related to ths performance fgure. From fgure 4, we note that the MD and L metrcs provde smlar performance. Ths appears motvated by the fact that performance are the trade off of two dfferent phenomenon. The MD metrc s targeted to optmze the channel qualty of each staton, regardless of the dstrbuton of MSs among cells. On the other sde, the L metrc s targeted to maxmze the spare capacty n each cell. However, although L balancng mproves the retransmsson effectveness, admtted calls suffer of worse channel condtons, snce they 894

5 are, n average, more dstant from the servng BS, wth respect to the MD selecton crteron. As shown n fgure 4, these two effects balance each other. Conversely, packet level metrcs are targeted to optmze, at the same tme, channel qualty and retransmsson effectveness. Therefore, ther performance are always superor wth respect to both MD and L metrcs. We recall that L balancng was consdered n the lterature to mnmze the call blockng probablty for new ncomng calls. Ths fact s shown n fgure 5, whch plots the call blockng rato versus the offered load for the four consdered metrcs. We see that our PL and GL metrcs have blockng performance that reman n the mddle between the worst case performance of MD and the best performance of L. We may argue that the performance comparson presented n fgure 4 s based by the fact that (wth a same offered load, on the bass of the results presented n fgure 5) the L metrc has a greater call acceptance rato, and thus t necessarly leads to more congested cells. Fgure 6 provdes a farer comparson, showng the packet loss rato versus the accepted load. Also n ths case, packet level metrcs prove ther clear-cut superorty wth respect of tradtonal metrcs. Fg. 7. Performance evaluaton wth correlated error model: normalzed offered load = 0.7, SR 0 /b=5,m=20 slots.fnally, fgure 7 shows the packet error rato when a correlated error model (as descrbed n secton III.A) s consdered. Clearly, the longer the channel error burst, the least lkely s that a faled packet transmsson can be recovered wthn a tme 1,E-05 1,E-06 1,E-07 PL L MD GL 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 Accepted Load Fg. 6. Packet loss versus normalzed cell throughput: SR 0 /b=5,m=20slots. PL MD L GL Average Length of an error burst frame (.e. for correlated error channels, retransmsson-based schedulng loses most of ts effectveness). Ths result s put nto lght n fgure 7, whch shows that, as long as the average error burst s greater than 8, all metrcs provde smlar performance (the excepton s the L metrc where each MS suffer, n average, of a transmsson error probablty greater than n the other metrcs). Indeed, for lower error burst szes the merts of packet level metrcs are confrmed. V. COCLUSIOS In ths paper, we have proposed two novel load balancng metrcs for wreless packet networks. The frst metrc, called Gross Load (GL), accounts for each connecton n terms of ts average channel slots consumpton per successful packet transmsson. A second, more complex, but slghtly more effectve, metrc, called ( PL), has been desgned to drectly estmate the packet loss performance experenced n a cell, and consequently allows moble statons to always attach to the cell for whch the best QoS s expected. The performance of the above metrcs has been compared wth that of tradtonal load balancng schemes. In all cases, the proposed approaches have been proven superor. Ths was clearly expected, as our schemes take advantage of addtonal packet level nformaton whch s not avalable n load balancng approaches based on call level nformaton only (e.g. the cell occupancy status n terms of number of admtted calls). REFERECES [1] B. Eklundh, Channel utlzaton and blockng probablty n a cellular moble telephone system wth drected retry, IEEE Trans. Commun., vol. 34, pp , [2] J. Karlsson, B. Eklundh, A cellular moble telephone system wth load sharng an enhancement of drected retry, IEEE Trans. Commun., vol. COM-37, pp , May [3] X. Lagrange, B. Jabbar, Farness n Wreless Mcrocellular etworks, IEEE Trans. on Vehcular Technology, vol. 47, no. 2, pp , May [4] T. Chu, S. Rappaport, Overlapng Coverage wth Reuse parttonng n Cellular Communcaton systems, IEEE Trans. on Vehcular Technology, vol. 46, no. 1, pp.41-54, Feb [5] G. Banch, F. Borgonovo, L. Fratta, L. Musumec, M. Zorz, ``C-PRMA: a Centralzed Packet Reservaton Multple Access for Local Wreless Communcatons'', IEEE Trans. on Vehcular Technology, May [6] F. Borgonovo, A. Capone, L. Fratta, "Retransmssons Versus FEC Plus Interleavng for Real-Tme Applcatons: A Comparson Between CDPA and MC-TDMA Cellular Systems", IEEE Journal on Selected Areas n Communcatons, vol. 17, no. 11, ov [7] D.Raychaudhur, L.J.French, R.J. Sracusa, S.K. Bswas, R. Yuan, I. arasmhan, C. A. Johnston, "WATMnet: a prototype wreless ATM system for multmeda personal communcatons", IEEE Journal of Selected Areas n Communcatons, vol. 15, no. 1, Jan. 1997, pp [8] M. J. Karol, Z. Lu, K. Y. Eng, "An Effcent Demand- Assgnment Multple Access Protocol for Wreless Packet (ATM) etworks), ACM Journal on Wreless etworkng, Dec. 895