VoIP Capacity over Multiple IEEE 802.11 WLANs *

VoIP Capacty over Mltple IEEE 802.11 WLANs * An Chan an Song Chang Lew Department of Informaton Engneerng The Chnese Unversty of Hong Kong Hong Kong, Chna {achan5, song}@e.chk.e.hk Abstract It s well known that IEEE 802.11 WLAN s hghly neffcent for transportng voce ata. For eample, f one smply takes the ata rate of 802.11b, 11Mbps, an ve t by two tmes 13.2Kbps (the bt rate of a typcal voce stream n one recton), one comes to the conclson that more than 400 voce sessons can be spporte n an 802.11b WLAN. As shown n prevos work, t trns ot that at most 12 sessons can be spporte e to varos heaer an protocol overheas nherent n 802.11. Ths paper ponts ot that the ba news oes not stop there, an that n practce the nmber of spportable voce sessons col be lower than 2 sessons per access pont (AP)! Ths s so becase as 802.11 WLAN gans poplarty, t s common to have many WLANs beng eploye n the same geographcal area, an these WLANs share the common ar mem. Or ns2 smlaton eperments, for eample, show that the capacty of a 5-by-5, 25-cell IEEE 802.11b WLAN, la ot n a sqare gr manner, s only 1.6 sessons per AP. The secon contrbton of ths paper s the nvestgaton of technqes to mprove the smal capacty. We show that a systematc call amsson mechansm base on clqe analyss of a conflct graph can ncrease the capacty to 2.12 sessons per AP. Ang a Restart Moe to the 802.11 protocol boosts the capacty frther to 2.72 sessons per AP. We also brefly scss the mpact of hgher ata rates (e.g., 54 Mbps n 802.11g an 802.11a), an carefl assgnment of avalable freqency channels (e.g., the three an twelve orthogonal freqency channels n 802.11b/g an 802.11a, respectvely), on capacty. Althogh all the above technqes can mprove the capacty somewhat, the hge penalty relatve to the potental remans. Boostng the voce capacty over mltple WLANs s therefore an area that eserves frther attenton from the research commnty. Keywors VoIP capacty; mltple WLANs; call amsson control; conflct graph; clqe I. INTRODUCTION Mch work [1-5] has been evote to the sty of VoIP (Voce-over-IP) capacty of IEEE 802.11 WLAN (Wreless LAN). The pror work focses prmarly on the sngle-cell envronment n whch there s only one solate WLAN wth one AP. Ths paper, n contrast, s the frst attempt to eamne the VoIP capacty n the mlt-cell envronment wth many WLANs beng eploye n the same geographcal area. Ths *Ths work was spporte by the Compettve Earmarke Research Grant (Project Nmber 414305) establshe ner the Unversty Grant Commttee of the Hong Kong Specal Amnstratve Regon, Chna. scenaro s common n practce. Pror nvestgaton has shown that even n the sngle-cell scenaro, the VoIP capacty over WLAN s severely lmte e to the varos nherent heaer an protocol overheas. Assmng the GSM 6.10 coec, for eample, only 12 VoIP sessons can be spporte n an 802.11b WLAN [3, 5]. Ths paper fns that the VoIP capacty s frther eroe n the mlt-cell scenaro, an sbstantally so. For eample, or ns2 [6] smlatons show that the capacty of a 5-by-5, 25-cell 802.11b WLAN s only 1.6 VoIP sessons per AP. Ths smal performance has mportant mplcatons that eserve frther attenton n vew of the acceleratng proctzaton of VoIPover-WLAN technologes. Althogh VoIP over WLAN has varos avantages sch as low cost, easy setp, ntegraton of voce an ata networks, an seamless moblty when combne wth celllar networks, ts actal performance n the real-worl envronment has not been well teste. Whle the performance can be acceptable f VoIP-over-WLAN remans npoplar an there are only a few sers, the poplarty of VoIP-over-WLAN can actally brng abot ts ownfall wth ts stretche capacty. The secon contrbton of ths paper s the nvestgaton of varos technqes to mprove VoIP capacty. Base on a conflct-graph moel, we set p a framework for call amsson. The smlaton reslts show that call amsson sng a clqe analyss can ncrease the capacty to 2.12 VoIP sessons per AP for the 5-by-5, 25-cell WLAN (.e., 32.5% mprovement). Another way to mprove capacty s to rece the conflct eges n the conflct graph by mofyng the MAC protocol. We show that a so-calle Recever Restart Moe, a mnor varaton of the 802.11 MAC protocol n some commercal chps, can ncrease the capacty of a 5-by-5, 25-cell WLAN by 50% to 2.4 VoIP sessons per AP. Combnng clqe-base call amsson an recever restart moe yels 2.72 sessons per AP for a 70% mprovement. In the later part of the paper, we also brefly scss the mpact of hgher ata rates (e.g., 54 Mbps n 802.11g an 802.11a), an carefl assgnment of avalable freqency channels (e.g., the three an twelve orthogonal freqency channels n 802.11b/g an 802.11a, respectvely), on capacty. Althogh all the above technqes can mprove the capacty somewhat, the hge penalty relatve to the potental remans. We therefore concle that boostng the voce capacty n

WLAN n the practcal settng s a frtfl area that eserves frther attenton from the research commnty. The prelmnary sty an the framework as set ot by ths paper serve as a goo startng pont for ong that. Althogh or nmercal stes assme the nfrastrctre-moe WLAN, the technqes an prncples emonstrate n ths paper can also be apple to a hoc networks. The remaner of ths paper s organze as follows. Secton II states the assmptons an network topologes se n or stes. The low capacty of VoIP over mltple WLANs s emonstrate. Secton III proves an analytcal eplanaton for the observaton n Secton II. Secton IV presents the reslts of the clqe-base amsson control an Recever Restart Moe. Use of 802.11g/802.11a an orthogonal freqency channels to ncrease capacty are scsse n Secton V. Secton VI concles the paper. II. VOIP OVER MULTIPLE WLANS A. Assmptons an Topology Settngs For concreteness, or stes assme the se of GSM 6.10. However, the general sses aresse n ths paper are the same for other coecs. We also assme the se of 802.11 Dstrbte Coornaton Fncton (DCF) [7] MAC protocol. Althogh Pont Coornaton Fncton (PCF) [7] s also worth styng, t s not poplar an almost never eploye n practce. Ths s perhaps becase () PCF s not spporte n most 802.11 procts an that () the robstness of PCF n real fel eployment has not been well teste, partclarly n statons where there are mltple mtally-nterferng APs n the vcnty of each other servng as the pont coornators. For or smlatons an analyses, we prmarly focs on 802.11b. In Secton V, we also brefly scss the se of 802.11a an 802.11g. For the GSM 6.10 coec, the voce payloa s 33-byte an 50 packets are generate n each secon. After ang the 40-byte IP/UDP/RTP heaer, the mnmm channel capacty to spport a voce stream n one recton (ether plnk or ownlnk) s 29.2Kbps. If we allow a packet-loss rate of 1% to 3%, the mnmm channel capacty reqrement s 28.32Kbps. We se ths mnmm channel capacty reqrement as a benchmark n the evalaton of the capacty of VoIP over WLAN. If both the plnk an ownlnk of a VoIP sesson can have throghpts eceeng ths benchmark, we say the VoIP sesson can be spporte n the WLAN. We moel a WLAN cell wth a sqare area of 353m 353m. An AP s place at the center of the cell. Any wreless clent staton nse the cell wll be assocate wth the AP. The longest lnk stance s 250m, whch s the ata transmsson range for 802.11b assme n ns2. By placng the cells se by se, we form a D-by-D mlt-cell topology, where D s the nmber of cells on each se. Fg. 1 shows a 2-by-2 mlt-cell topology. B. 2-by-2 Mlt-Cell Scenaro We frst conser the small-scale 2-by-2 mlt-cell topology n Fg. 1. Clent statons (VoIP sessons) are ae to the network one by one ranomly assmng nform strbton. Wth each atonal VoIP sesson, ns2 smlaton s rn an the throghpt of each lnk recore. When the net newly ae VoIP sesson cases a volaton of the packet-loss rate reqrement by at least one of the sessons, we say that the capacty lmt has been eceee. Ths correspons to a smplstc call amsson scheme n whch pon the above nacceptable performance case by the newly ae sesson, the newly sesson wll be roppe, an no more ftre sessons wll be accepte. The reason no more ftre sessons wll be accepte s smply becase the ang an roppng of QoS-volatng sessons can be qte srptve to the qalty of the estng sessons. We wll later conser a cleverer call amsson scheme base on a clqe analyss of a conflct graph so that we can prect the performance before ecng whether to amt a call to get r of the srptveness. Or smlaton reslts ncate that at most 12 sessons can be spporte n the overall 2-by-2 network. Havng for APs s the same as havng one AP as far as the overall capacty s concerne! The nmber of VoIP sessons per AP s only three. C. 5-by-5 Mlt-Cell Scenaro We now look at a larger-scale 5-by-5, 25-cell scenaro. Smlar to the smlatons n 2-by-2 case, we progressvely a more clents to the topology ntl the net VoIP sesson cases volaton of packet-loss rate reqrement by at least one of the sessons. We have performe several rns of ns2 smlatons, each wth fferent clent-staton strbtons on the 5-by-5 mltcell topology. The nmber of VoIP sessons spporte ranges from 38 to 47, wth an average of 40. Fg. 2 shows the reslt of a typcal rn when there are 40, 41, 42 VoIP sessons. We recor the evaton of the throghpt of each lnk from the 28.32Kbps benchmark (Devaton = Throghpt Benchmark). In Fg. 2, the legen 40-DL means ownlnk throghpts when there are 40 VoIP sessons. UL means plnk throghpts. We see that when there are 40 VoIP sessons (sol bol lnes), all the lnks have throghpts eceeng the benchmark (postve evaton vales). Bt when the nmber of VoIP sessons ncreases to 41 or 42 (otte lnes), some lnks have throghpts lower than the benchmark (negatve evaton vales). The VoIP capacty n ths smlaton rn s therefore 40 sessons. 706m 706m Magnfe vew 250m 353m Fg. 1. A 22 mlt-cell topology AP

Devaton from Benchmark (Kbps) 1 0.5 0-0.5 40-DL 40-UL 41-DL 41-UL 42-DL 42-UL In the sngle-cell scenaro, all clent statons are wthn the PCS range of each other becase they are assocate wth the same AP. So, eges shol be rawn among all vertces. That means any newly ae verte wll ncrease the clqe sze by one. The analyss an smlatons n [5] showe that the mamm nmber of VoIP sessons (assmng GSM 6.10 coec) that can be spporte n an 802.11b sngle-cell WLAN s 12, becase 12 VoIP sessons wll fll p all artmes an no artme s left for an atonal VoIP sesson. Therefore, 12 s also the mamm sze of a clqe n the conflct graph. 0 3 6 9 12 15 18 21 24 27 30 33 36 39 Lnk ID Fg. 2. Devaton of throghpt of each lnk from benchmark when there are 40, 41 an 42 VoIP sessons n a 55 WLAN The nmber of VoIP sessons that can be spporte per AP s aron 1.6 (40/25), whch s less than two per AP! Compare wth the sngle-cell scenaro, where each AP can spport 12 VoIP sessons, ths s a rather large penalty. We also note n passng that for a network larger than the 5-by-5 network, most cells wll be srrone by eght ajacent cells an there wll be proportonately fewer cells at the bonary. One can therefore epect the VoIP capacty per AP to rop even frther when the mensons are larger than 5-by-5. III. CLIQUE ANALYSIS To nerstan the case for the heavy performance penalty, we conser here a clqe analyss base on a graph moel that captres the conflct an nterference among the noes. The clqe analyss also sggests a call amsson methoology later. A. Conflct Graph an Clqes In a conflct graph, vertces represent VoIP sessons. An ege between two vertces means that the two VoIP sessons compete for the artme. In other wors, they cannot transmt at the same tme. There are two reasons why they cannot transmt at the same tme. () Frst, noes of the two sessons are wthn the Physcal Carrer Sensng (PCS) range (note: we conser only the basc access moe n ths paper) of each other wll be prevente by the 802.11 protocol from transmttng together. () Even f the two sessons are not wthn each other s PCS range, there may be mtal nterference between them so that ether one or both of ther transmssons wll fal f they transmt together. Ths s e to the wellknown hen-noe problem [8]. In ether case () or (), an ege s rawn between the two vertces of the VoIP sessons to mean that ther sccessfl transmssons cannot share the same artme. As more VoIP sessons are establshe, more vertces an eges wll be ntroce n the graph. A clqe s a sbset of vertces wth an ege between any of two of them. The vertces n a clqe compete for the common artme. That s, the sm of the fractons of artmes se by the vertces shol not ecee one. B. Clqe n 2-by-2 Mlt-Cell Topology Wth reference to Fg. 1, we fn that the farthest stance between two APs n the 2-by-2 mlt-cell topology s 500m. Snce the assme PCS range s 550m, all for APs are wthn the PCS range of each other. The AP of the cell mst be a recever or a transmtter of any VoIP sesson nse the cell, so once for APs are covere by the PCS range of each other, all VoIP sessons n the 2-by-2 mlt-cell topology are covere by the PCS range of each other. Hence, there s an ege between any par of vertces nse the corresponng conflct graph of the 2-by-2 mlt-cell topology. Ths bols own to the same staton as the sngle-cell case, an that the mamm clqe sze s 12. Ths s consstent wth the smlaton reslt we presente n the prevos secton an eplans the low capacty of the 2-by-2 mlt-cell WLAN. C. Clqes n 5-by-5 Mlt-Cell Topology The sze of the 5-by-5 mlt-cell topology s larger. Some VoIP sessons are otse the PCS range (case () n sbsecton A) an nterference range (case () n sbsecton A) of each other. That means we can fn pars of vertces between whch there s no ege, an therefore ther sccessfl transmssons can share the same artme. Fg. 3 shows an eample of sch VoIP sessons. The nearest stance between the for shae cells s 706m, an therefore the noes nse fferent shae cells are not covere by each other s PCS range of 550m, as s assme n ns2. In other wors, case () n the frst paragraph of sbsecton A s not satsfe. For case (), we conser the Interference Range (IR) efne as follows: IR = (1 + ) (1) where IR of a noe (can be a clent staton or an AP) s the 706m 1765m 706m Fg. 3. A 55 mlt-cell WLAN topology 1765m

range wthn whch any other transmsson wll nterfere wth the transmsson to or from the noe; s the length of the lnk assocate wth the noe; an s a stance margn for nterference-free recepton wth typcal vale of 0.78 [6, 8]. The mamm lnk length wthn a cell n Fg. 3 s ma = 250m (shown n Fg. 1). The corresponng mamm IR s IR ma = 1.78 250 = 445m As shown n Fg. 3, the stance between any two ponts of two fferent shae areas s beyon more than IR. Therefore, there s no ege between ma vertces of fferent shae cells, an a VoIP sesson n one shae cell can transmt smltaneosly wth a sesson n another shae cell. Snce some verte pars n the conflct graph of the 5-by-5 mlt-cell WLAN o not have an ege, two or more clqes wth the mamm sze can potentally be forme. Ths eplans why the total nmber of VoIP sessons that can be spporte n the 5-by-5 mlt-cell WLAN ecees 12 (recall that smlaton reslts n Secton II show 40 sessons can be spporte on average). Note, however, that the nmber of cells covere by the PCS range of a cell n the 5-by-5 mlt-cell WLAN s larger than that n the overall area of the 2-by-2 mlt-cell WLAN (n the 2-by-2 mlt-cell WLAN, the PCS range of a cell actally etens beyon the whole topology). Ths then cases the nmber of VoIP sessons spporte per AP n the 5-by-5 case to be even lower than the 2-by-2 case (40/25=1.6 verss 12/4=3 sessons per AP). In 5-by-5 mlt-cell WLAN, there are 16 bonary-cells (the cells on the bonares of the topology) whch eperence less nterference from ether others voce packet transmsson or PCS range than the nne non-bonary-cells. The APs n the non-bonary-cells spport fewer VoIP sessons than those n the bonary-cells. As mentone n Secton II, for a network larger than the 5-by-5 network, most cells wll be srrone by eght ajacent cells an the rato of the nmber of bonary-cells to the nmber of non-bonary-cells wll ecrease. So, the average nmber of VoIP sessons that can be spporte by each AP wll rop even frther when the mensons are larger than 5-by-5. Or man fnngs n ths paper s that even wth 2-by-2 an 5-by-5 mlt-cell topologes, the average nmber of VoIP sessons spporte by each AP s alreay qte low. Wth a larger topology, thngs can only get worse. IV. METHODS TO INCREASE CAPACITY The analyss n the prevos secton shows that the mamm clqe sze shol not ecee 12. In or prevos smlatons, VoIP sessons are ae ntl the net VoIP sesson cases a volaton of the packet-loss rate reqrement by at least one of the sessons. A more jcos call amsson process s calle for. In other wors, even thogh sch a net VoIP sesson may case volaton of the lossrate reqrement, t oes not mean the call reqest after that net call wll also case a volaton. In partclar, we nee a more sophstcate amsson control algorthm whch amts VoIP sessons base on clqe sze. Ths may potentally n- crease the mamm nmber of spportable VoIP sessons. Another menson to ncrease the VoIP capacty s to rece the formaton of eges n the conflct graph, so that more VoIP sessons can be amtte before the mamm clqe sze s reache. The followng sbsectons scss these two mensons an propose possble soltons. A. Call Amsson Control Base on Clqe Sze The ablty to prect whether a new VoIP sesson can be amtte wthot casng performance problem s mportant n call amsson. Prevos work [2, 9] shows that an atonal VoIP sesson to a sngle-cell WLAN whch alreay reache the mamm capacty can egrae the performance of all other estng VoIP sessons. Ths s also the case n mlt-cell WLANs. In the 2-by-2 network, for eample, f there are 14 VoIP sessons, three of them cannot flfll the loss-rate reqrement. That means only 11 sessons can be spporte wth acceptable performance, as oppose to the 12 sessons that can be spporte when there are eactly 12 sessons n WLAN. In ths secton, we conser a call amsson control mechansm base on the clqe sze of the conflct graph to psh the se of VoIP capacty to a mamm, bt not more than that so as not to volate loss-rate reqrements. Let E be the set of vertces wth whch verte V has an ege. Let K be the set of all clqes C to whch V, V belongs, where 1 K V, K s the total nmber of clqes n K. A clqe C tself s a set contanng ts V, assocate vertces. Any par of clqes n K mst satsfy (2) below so that all clqes n K are mamal an not contane n another clqe. As efne n (3), m s the sze of the largest clqe n K. C, C, V, y y, 1, y K V (2) m = ma C (3) V V, Fg. 4 gves an eample of a conflct graph, where verte V has the followng parameters. 1 E = { V, V, V, V } V1 2 3 4 5 K = { VV,, VV, },{ VVV,, }, V1 1 2 3 5 1 3 4.e., K = 2 C = { V, V, V, V }, C = { V, V, V } V1,1 1 2 3 5 V1,2 1 3 4 4 m = V 4 V 3 V 5 Fg. 4. An eample of a conflct graph

The general problem of fnng clqes n a graph s NP-Complete. However, the conflct graph forme by VoIP sessons n a mlt-cell WLANs has certan propertes carre over from the physcal propertes of the contenton an nterference relatonshps among lnks. For eample, only the vertces (VoIP sessons) that are physcally close to each other have the possblty to form clqes. Or call amsson algorthm below tres to eplot these physcal propertes to spee p the eecton tme of the algorthm. The pseo coe of the amsson control algorthm s gven n Algorthm I. There are three proceres n the algorthm: Proceres A, B an C. When there s a new call reqest (.e., a new verte (VoIP sesson) wants to be ae), Procere A s frst eecte, wheren a copy of the state ( K, m ) s frst save n case the amsson of Vj Vj j V fals an we nee to revert to the orgnal state later. After that, Procere B s eecte. Procere B pates the state Algorthm I Procere A Keep a copy of the state ( KV, m ) for all estng V ; j Vj j Perform Procere B; Procere B For each Vj E { For each CVj, K { Vj If CVj, EV then a V to C ; Vj, else a { CV,, } j EV V to } K = NO_REDUNDANCY ( K ); Upate m ; If m > C, ma then reject V ; K ; Revert the state sng the copy store n Procere A; Break ot of Procere B, an the algorthm s termnate; } // All Vj E have been looke at f the algorthm comes to ths pont Perform Procere C; Procere C s amtte; K V = ; For each Vj E { If V CVj, then C, = C, Vj, KV = K V C ; V, } K = NO_REDUNDANCY ( K ); Compte m ; ( KV, m ) assmng the aton of j V V j. To satsfy (2), a fncton call NO_REDUNDANCY ( K ) s mae. Algorthm II gves the pseo coe of NO_REDUNDANCY ( K V k ). Drng the patng, Procere B contnally checks to see f a preetermne mamm clqe sze C ma s eceee so as not to volate the loss-rate reqrement. If so, the algorthm s termnate an s rejecte; n whch case the state save n Procere A s restore. If Procere B sccessflly rns to the en wthot the C ma beng eceee, Procere C s eecte. Procere C amts the new verte V an calclates ( K, m ). We have performe an eperment n MATLAB to measre the tme the call amsson control algorthm nees to accept or reject one call. In the eperment, we se the algorthm to fn whch VoIP sessons ot of a total of N nformly ranomly place potental sessons can be amtte n a 5-by-5 mlt-cell WLAN topology wth the C ma constrant on the largest clqe sze. Unlke the smplstc call amsson scheme earler, here when a sesson s rejecte, the call amsson scheme contnes to conser a net sesson f all the N sessons have not been ehaste. In the eperment, fferent N an fferent noe strbtons are consere n a 5-by-5 WLAN. We se an ornary personal compter wth 3.2GHz CPU an 1G RAM to perform the eperment. The reslts for fferent N an C ma are shown n Table I. The total rntme s the total tme neee for the amsson control algorthm to go throgh all the N VoIP sessons. For a gven C ma, the total rntme s rectly proportonal to N. The average rntme the algorthm neee to amt or reject a call (Total rntme / N) s 0.23s when C ma s 9, an 0.37s when C ma s 12. We see that althogh the general clqe problem s NP-complete, the eecton tme of or algorthm on the Algorthm II NO_REDUNDANCY ( K V k ){ For each par CV,, k KV C k Vk, y K, V y { k If CV k, CV k, y then eletec ; Vk, } Retrn ( K V k ); } TABLE I EXPERIMENT RESULTS OF THE ADMISSION CONTROL ALGORITHM WHEN DIFFERENT N AND C ma ARE USED N C ma Total (s) rntme No. of amtte sessons 75 9 18 52 12 31 64 100 9 20 52 12 37 68 125 9 28 53 12 47 71 150 9 36 58 12 51 76 V V

conflct graph that moels 802.11 networks s actally acceptable. Base on the call amsson reslts n MATLAB, we then se ns2 to verfy whether the amtte calls meet the mamm 3% packet-loss rate reqrement n the smlaton. From Table I, we fn that when C ma s 9, the nmbers of VoIP sessons amtte by or amsson control algorthm are 52, 52, 53, 58, for N = 75, 100, 125, 150, respectvely. We ncorporate the assocate for sets of amtte VoIP sessons n for fferent rns of ns2 smlaton. In each rn, all the amtte VoIP sessons can meet or packet-loss rate reqrement. Therefore, the average nmber of VoIP sessons meetng the packet-loss rate reqrement ner the call amsson control s aron 53. However, f we nstea se the for sets of the amtte VoIP sessons when C ma s 12 n the ns2 smlatons, nearly one thr of them cannot meet the loss reqrement. It s nterestng that for the large-scale mlt-cell case, the mamm clqe sze that shol be mpose s 9 rather than 12 (recall that 12 s the mamm clqe sze n the sngle-cell topology) f loss-rate reqrement s to be satsfe. Ths s perhaps e to the nteracton an coplng among fferent clqes case by the 802.11 MAC protocol. In other wors, 802.11 MAC may not acheve perfect schelng n whch the artme sage wthn each clqe s 100% tghtly packe. Wth or call amsson control above, the 5-by-5 mltcell WLAN can spport 2.12 sessons (53/25) per AP, yelng a 32.5% mprovement over the smplstc call amsson scheme n Secton II. B. Recever Restart Moe We now conser the other menson to sqeeze n more VoIP capacty: recng the eges among the vertces n the conflct graph. Recever Restart (RS) Moe [8] s ncorporate nto some commercal 802.11 chps (e.g. Atheros Chp). Wth RS, the recever swtches to receve a new sgnal n the mst of recevng an ol sgnal f the new sgnal has a power stronger than the ol sgnal by a certan threshol, say 10B. Ths moe changes the nteracton among lnks an may rece the eges n the conflct graph. To nerstan an moel the effect of RS Moe, we nee to conser the two transmsson rectons of a VoIP sesson separately. Let s splt each verte n the conflct graph nto two vertces, V an V, representng the plnk (clent staton to AP) an ownlnk (AP to clent) of separately (as shown n Fg. 5a an 5b, the etals of whch are eplane n the net few paragraphs). We shall refer to ths conflctgraph moel as recton-aware conflct graph moel, an the prevos conflct-graph moel as recton-oblvos conflct graph moel. We llstrate the ea n Fg. 5a an 5b. In the fgres, we assme the stances between noes satsfy (4) an (5). ( 1 + ) < (4) 1 3 ( 1+ ) 2 < (5) 3 where s the stance margn for nterference-free recepton. Accorng to [8], once (4) an (5) are satsfe, the two VoIP sessons ( an ) wll not nterfere wth each other. Bt ther PCS ranges may stll affect each other becase the PCS range n 802.11 s a fe vale that oes not change wth lnk stance. In Fg. 5a, the PCS range of N1 covers AP2, so an ege shol be rawn between an n the rectonoblvos conflct graph. In the recton-aware graph, ths ege may be mappe to a mamm of s eges. Frst, there s always an ege between an, an an ege between V an 2, respectvely, by the assmpton of half-plety. Then, there are the for potental eges between the par V an 1, an the par V an 2. Conser the nteracton between an. Sppose that N1 sens a packet to AP1 frst, followe by N2 senng a packet to AP2 (note that N2 can sen a packet becase t s not covere by the PCS range of N1). If the RS Moe s not trne on, then e to the coverage of the PCS range of N1, AP2 wll not respon to the packet from N2 even thogh two VoIP sessons actally can operate at the same tme accorng to the relatonshps n (4) an (5). The commercal 802.11 proct commonly operates wthot RS by efalt althogh some chps (e.g. Atheros Chp) have ncorporate RS as an optonal moe. The efalt operaton moe n ns2 s also wthot RS. Accorngly, there shol be an ege between V an wthot RS moe. 1 N1 3 1 2 AP1 PCS Range Fg. 5a. An eample to show the effect of RS Moe AP2 N2 On the other han, wth RS, AP2 wll swtch to receve the new sgnal from N2 an respon wth an ACK. That s, there shol be no ege between an n the rectonaware conflct graph. Therefore, wth RS, the nmber of eges s rece an more VoIP sessons can be potentally amtte to the WLAN before C ma s reache. For the recton-aware conflct graph, C ma shol be change a larger vale, say 18 (.e., two tmes the vale of C ma n the recton-oblvos conflct graph moel). In ns2 smlatons, we fn that a total of 60 VoIP sessons can be spporte n a 5-by-5 mlt-cell WLAN wth RS. That means each AP can spport 2.4 VoIP sessons, yelng a 50% mprovement Fg. 5b. The corresponng recton-aware conflct graph

relatve to the case wthot RS. If the call amsson control escrbe n the prevos sbsecton s se together wth RS, there s another 13.3% mprovement so that 2.72 VoIP sessons per AP can be spporte. V. USE OF 802.11g/a AND ORTHOGONAL FREQUENCY CHANNEL ASSIGNMENT TO BOOST CAPACITY Ths paper has so far assme 802.11b. In ths secton, we brefly scss the se of 802.11a an 802.11g to boost VoIP capacty. In aton, we also prove a qck assessment of assgnng orthogonal freqency channels to ajacent cells n mlt-cell WLANs. A. Stanars wth Hgh Data Rates In the IEEE 802.11 stanar famly, 802.11a an 802.11g can offer hgher ata rates (p to 54Mbps) than 802.11b. So, more VoIP sessons can be spporte n 802.11a or g WLANs. From the overhea analyss n [5], the mamm nmber of VoIP sessons s 56 n a sngle-cell 802.11a WLAN, an 60 n 802.11g sngle-cell WLAN. However, the performance penalty of the mlt-cell case relatve to the sngle-cell case s e to the nterference from neghborng cells an the 802.11 carrer-sensng mechansm, whch est n all 802.11 stanars regarless of the ata rate se. Ths, the penalty for 802.11a or g when we go from the sngle-cell staton to the mlt-cell staton s smlar to that n 802.11b, althogh the absolte nmber of VoIP sessons that can be spporte s hgher. In partclar, base on the reslts of 802.11b, we prect that aron 8 an 9 VoIP sessons per AP can be spporte by 802.11a an 802.11g, respectvely, n a 5-by-5 mlt-cell WLAN. That s, the nmber of VoIP sessons that can be spporte per AP n a mlt-cell WLAN s stll small even thogh the hgher ata rate s se. Frthermore, the coverage area of the AP for the mamm ata rate of 54Mbps s small, an for larger coverage areas, lower ata rates mst be se. Ths, we are stck n a lemma n whch we may want to se a large coverage area to serve more sers, bt the VoIP capacty for the larger coverage area can be sgnfcantly lower. B. Channel Assgnment n Dfferent Cells The nterference from neghborng cells can be solate f fferent freqency channels are se n ajacent cells. If there are enogh freqency channels, the neghbor cells that may nteract wth a cell throgh nterference an carrer sensng col be assgne fferent freqency channels. Ths bols own to the same staton as the sngle-cell case, so the per-ap VoIP capacty n the mlt-cell case s the same as that n the sngle-cell case. However, 802.11b an 802.11g have only three orthogonal freqency channels, an ths s not enogh to completely solate co-channel nterference between cells. For channel assgnment, we col se heagonal cells for layng ot the mltple cells. Fg. 6 shows that f we have only three freqency channels, then the nearest stance between two cells sng the same channel s the same as the mamm lnk length wthn a cell, ma. Snce these two cells may nterfere wth each other, the carrer-sensng range (CS range) shol be larger than 3 ma to avo hen-noe collsons between the two cells (conser lnk (AP1, STA1) an lnk (AP2, STA2) of the two cells, where the stance between STA1 an STA2 s ma ; carrer-sensng range of 3 ma s neee to avo hen-noe collsons between these two lnks). Snce a cell mst now share artme wth other cells, the VoIP capacty per AP cannot be the same as that n the sngle-cell case. 802.11a, on the other han, has twelve orthogonal channels. Fg. 7 shows that a 7-channel assgnment s enogh for complete solaton of co-channel nterference. The nearest stance between two cells sng the same channel s 2.65 ma whch s larger than the mnmm CS range, 2 ma, se to prevent collsons wthn a cell. So wth 7-channel assgnment, co-channel nterference can be completely solate. However, f we smply se 7 overlay networks n each cell (pt 7 APs nse each cell an operate n fferent channels), the nmber of VoIP sessons that can be spporte n each cell s aron 56 (78, snce we prect that 8 sessons can be spporte per AP n 802.11a mlt-cell WLAN n prevos sbsecton). Ths s the same as the capacty of 802.11a sngle-cell WLAN. If RS Moe an call amsson control are se n the overlay networks, the capacty can be hgher than 56 VoIP sessons per cell. Therefore, we fn that the channel assgnment n Fg. 7 actally oes not mprove the VoIP capacty on a per-cell (or per-freqency channel) bass, althogh on a per-ap bass, t oes. VI. CONCLUSION Ths paper has eamne the VoIP capacty of IEEE 802.11 mlt-cell WLAN. The VoIP capacty of the sngle-cell case, 12 sessons per AP for 802.11b, s alreay qte low e to varos protocol an heaer overheas. There s another large capacty penalty when we move to the mlt-cell case: only an average of 2.12 sessons per AP for a 5-by-5, 25-cell network. An ths can be acheve only wth a carefl call amsson scheme base on a conflct-graph clqe analyss; ma AP ma Fg. 6. 3-channel assgnment n mlt-cell WLAN 2.65 ma Fg. 7. 7-channel assgnment n mlt-cell WLAN

the average capacty s only 1.6 sessons per AP wth a smplstc call amsson. The conflct-graph clqe analyss consere n ths paper, beses beng closely te to or call amsson scheme, also yels nsghts on how VoIP capacty can be ncrease. In partclar, we col conser mofyng the 802.11 MAC protocol an carrer-sensng mechansm to rece the nmber of eges n the conflct graph, hence recng the clqe sze. Ths paper has consere a so-calle RS moe to rece eges. RS moe only nees a mnor mofcaton of the 802.11 DCF protocol, an can ncrease the VoIP capacty by 50%. Combnng RS wth the clqe-base call amsson scheme yels a capacty of 2.72 sessons per AP n the 5-by-5, 25-cell network. It s a 70% capacty mprovement relatve to the case of sng smplstc call amsson wthot RS. Even f RS moe an clqe-base call amsson are se together, the VoIP capacty s stll qte low. There s mch room for mprovement. We have arge that althogh 802.11a an g col rase the absolte VoIP capacty e to ther hgher ata rates, the relatve large capacty penalty as compare to the sngle-cell case remans (.e., the capacty rato of the mlt-cell an sngle-call cases remans largely the same). Ths paper has also attempte a prelmnary foray nto the sse of channel assgnment: how best to se the three orthogonal freqency channels n 802.11b/g an the twelve orthogonal freqency channels n 802.11a. Or conclson s that althogh the VoIP capacty on a per-ap bass can ncrease somewhat, the VoIP capacty on a per-freqency s stll low. A more etale analyss awats frther work. Frther nvestgatons to boost VoIP capacty col procee along the rectons of () recton of heaer an protocol overheas by the se of packet aggregaton methos [3], or brst transmsson moe n 802.11e; () recton of conflct eges by means of avance 802.11-lke protocols [10,11], power control [12,13], or rectonal antennas [14]; an () call assgnment schemes n whch sessons n hotspot cells are assgne an assocate to ajacent cells [15]. Approach () s relevant to both sngle- an mlt-cell scenaros whle approaches () an () are nqe to the mlt-cell scenaro. Technology, vol. 54, no. 1, Jan. 2005. [6] The Network Smlator ns2, http://www.s.e/nsnam/ns/ [7] M. S. Gast, 802.11 Wreless Networks: Defntve Ge, O Relly, Sebastopol, CA, 2002. [8] L. Jang, S. C. Lew, Removng hen noes n IEEE 802.11 wreless networks, IEEE Vehclar Technology Conference, Sep. 2005. More comprehensve verson to appear as Hen-noe removal an ts applcaton n celllar WF networks, IEEE Trans. On Vehclar Technology, Nov 2007. [9] L. Ca, X. Shen, J. W. Mark, L. Ca, Y. Xan, Voce capacty analyss of WLAN wth mbalance traffc, Proceengs of Qshne 05, Ag. 2005 [10] P. C. Ng, S. C. Lew, an L. Jang, Achevng scalable performance n large-scale IEEE 802.11 wreless networks, IEEE Wreless Commncatons an Network Conference (WCNC), Mar 2005 [11] A. Chan an S. C. Lew, Mert of PHY-MAC cross-layer carrer sensng: a MAC-aress-base Physcal Carrer Sensng scheme for solvng hen-noe an epose-noe problems n large-scale W-F networks, The 6 th IEEE Workshop on Wreless Local Networks, Nov 2006. [12] W. H. Ho an S. C. Lew, Achevng scalable capacty n wreless networks wth aaptve power control, The 5 th IEEE Workshop on Wreless Local Networks, Nov 2005. [13] W. H. Ho an S. C. Lew, Dstrbte power control n IEEE 802.11 wreless networks, IEEE Internatonal Conference on Moble A-hoc an Sensor System (MASS), Oct 2006. [14] J. Zhang an S. C. Lew, Capacty mprovement of wreless a hoc networks wth rectonal antennae, IEEE Vehclar Technology Conference, May 2006. [15] S. C. Lew an Y. J. Zhang, Proportonal farness n mlt-channel mlt-rate wreless networks, IEEE Globecom 2006, Nov 2006. REFERENCES [1] F. Anjm et al, Voce performance n WLAN networks an epermental sty, Proceengs of IEEE Globecom 03, vol. 6, pp. 3504-3508, Dec. 2003. [2] S. Garg, M. Kappes, An epermental sty of throghpt for UDP an VoIP traffc n IEEE 802.11b networks, Proceengs of IEEE WCNC 03, vol. 3, pp. 1748-1753, Mar. 2003. [3] W. Wang, S. C. Lew, Q. X. Pang, an V. O.K. L, A mltple-mltcast scheme that mproves system capacty of Voce-over-IP on wreless LAN by 100%, The Nnth IEEE Symposm on Compters an Commncatons, Jne 2004 [4] D. P. Hole, F. A. Tobag, Capacty of an IEEE 802.11b wreless LAN spportng VoIP, Proceeng of the IEEE ICC 04, vol. 1, pp. 196-201, Jn. 2004. [5] W. Wang, S. C. Lew, V. O. K. L, Soltons to performance problems n VoIP over 802.11 wreless LAN, IEEE Trans. on Vehclar