VoIP over Multiple IEEE 802.11 Wireless LANs

SUBMITTED TO IEEE TRANSACTIONS ON MOBILE COMPUTING 1 VoIP over Multple IEEE 80.11 Wreless LANs An Chan, Graduate Student Member, IEEE, Soung Chang Lew, Senor Member, IEEE Abstract IEEE 80.11 WLAN has hgh data rate (e.g., 11Mbps for 80.11b and 54Mbps for 80.11g) whle voce stream of VoIP has low data rate requrement (e.g., 9.Kbps). One may therefore expect WLAN to be able to support a large number of VoIP sessons (e.g., 00 and 900 sessons n 80.11b and 80.11g, respectvely). Pror work by one of the authors, however, ndcated that 80.11 s extremely neffcent for VoIP transport. Only 1 and 60 VoIP sessons can be supported n an 80.11b and an 80.11g WLAN, respectvely. Ths paper shows that the bad news does not stop there. When there are multple WLANs n the vcnty of each other a common stuaton these days the already-low VoIP capacty can be further eroded n a sgnfcant manner. For example, n a 5-by-5, 5-cell mult-wlan network, the VoIP capactes for 80.11b and 80.11g are only 1.63 and 10.34 sessons per AP, respectvely. Ths paper nvestgates several solutons to mprove the VoIP capacty. Based on a conflct graph model, we propose a clque-analytcal call-admsson scheme, whch ncreases the VoIP capacty by 5% from 1.63 to.48 sessons per AP n 80.11b. For 11g, the call-admsson scheme can also ncrease the capacty by 37%, from 10.34 to 14.14 sessons per AP. If all the three orthogonal frequency channels avalable n 11b and 11g are used to reduce nterferences among adjacent WLANs, clque-analytcal call admsson can boost the capacty to 7.39 VoIP sessons per AP n 11b and to 44.91 sessons per AP n 11g. Last but not least, ths paper expounds for the frst tme the use of coarse-graned tme-dvson multple access (CoTDMA) n conjuncton wth the basc 80.11 CSMA to elmnate the performance-degradng exposed-node and hdden-node problems n 80.11. A -layer colorng problem (whch s dstnct from the classcal graph colorng problem) s formulated to assgn coarse tme slots and frequency channels to VoIP sessons, takng nto account the ntrcaces of the carrer-sensng operaton of 80.11. We fnd that CoTDMA can further ncrease the VoIP capacty n the mult- WLAN scenaro by an addtonal 35% to 10 and 58 sessons per AP n 80.11b and 80.11g, respectvely. Index Terms VoIP, multple WLANs, CSMA, coarse-graned tme-dvson multple access, clque-analytcal call admsson control 1 INTRODUCTION V OICE-OVER-IP (VoIP) s one of the fastest growng applcatons for the Internet today. At the same tme, drven by huge demands for portable access, the market for wreless Local Area Network (WLAN) based on the IEEE 80.11 standard s takng off quckly. Many ctes are plannng on cty-wde deployment of WLAN. An mportant applcaton over these networks wll be VoIP over WLAN. A hurdle, however, s the low number of voce conversatons that can be supported. As shown n prevous nvestgatons [1, ] by one of the authors, although n theory many voce sessons can be supported n an 80.11b WLAN based on smplstc raw-bandwdth calculaton, n realty only less than 1 can be accommodated. There has been much subsequent work on how to mprove the qualty-of-servce (QoS) of VoIP over WLAN. Part of the IEEE 80.11e standard [3], for example, s to address that. Most of the pror nvestgatve efforts [1,, 4-6], have been focused on the sngle solated WLAN scenaro. In A. Chan graduated from the Department of Informaton Engneerng, The Chnese Unversty of Hong Kong. He s now studyng Ph.D program n the Department of Computer Scnece n the Unversty of Calforna, Davs, CA 95616. E-mal: achan5@ e.cuhk.edu.hk. S.C. Lew s wth the Department of Informaton Engneerng, The Chnese Unversty of Hong Kong, Hong Kong. E-mal: soung@e.cuhk.edu.hk. Manuscrpt receved December 1, 007. practce, wth the prolferaton of WLAN these days, t s common to fnd numerous WLANs wthn a small geographcal area one only needs to do a cursory scan wth a WLAN-equpped personal computer to see the consderable number of WLANs wthn a buldng. Although recently there has been much attenton pad to multhop wreless mesh networks and VoIP over such networks [7-9], nfrastructure WLAN s stll the most wdely deployed archtecture nowadays. Ths paper s a frst attempt to examne the VoIP capacty n the mult-cell envronment n whch many WLANs are deployed n the same geographcal area. We fnd that the VoIP capacty s further eroded n the mult-cell scenaro, and substantally so. For example, our NS [10] smulatons show that the capacty of a 5-by-5, 5-cell WLAN s only 1.63 VoIP sessons per access pont (AP) n 80.11b, and 10.34 sessons per AP n 80.11g. Ths dsmal performance has mportant mplcatons that deserve further attenton n vew of the acceleratng productzaton of VoIP-over-WLAN technologes. Besdes pontng out the alarmngly low effcency of VoIP over mult-cell WLAN, and dentfyng the mutual nterferences of the CSMA operaton of adjacent cells as the major culprt for the dsmal performance, ths paper s also a frst foray nto fndng solutons to allevate the problem. Based on a conflct-graph model, we set up a framework for call admsson control so as to better manxxxx-xxxx/0x/$xx.00 00x IEEE

SUBMITTED TO IEEE TRANSACTIONS ON MOBILE COMPUTING age the mutual nterferences. The smulaton results show that a clque-analytcal call-admsson scheme can ncrease the VoIP capacty to.48 VoIP sessons per AP n 80.11b (.e., 5.1% mprovement) and to 14.14 sessons per AP n 80.11g (.e., 36.75% mprovement) for the 5-by- 5, 5-cell WLAN. If all the three orthogonal frequency channels n 80.11b/g are used, the clque-analytcal call admsson control can boost the per-ap capacty to 7.39 VoIP sessons n 80.11b and to 44.91 sessons n 80.11g. Another major contrbuton of ths paper s the proposal of a coarse-graned tme-dvson multple-access (CoTDMA) approach to allevate the mult-cell mutual nterferences. In CoTDMA, the tme dmenson s dvded nto multple coarse tme slots. Multple VoIP sessons are then assgned to each tme slot, and they make use of the basc 80.11 CSMA protocol to coordnate channel access wthn the tme slot. Coarse-graned tme slots could be mplemented usng the sleep mode of 80.11, orgnally ntended for power conservaton purposes. The basc dea of CoTDMA s that VoIP sessons of adjacent WLANs that nterfere wth each other should be assgned dfferent tme slots, so that VoIP sessons of dfferent cells do not need to use CSMA to coordnate transmssons among them; essentally CSMA needs to be effectve only among the sessons of the same cell. As wll be shown n ths paper, the theoretcal calladmsson control framework of CoTDMA corresponds to a new class of graph-colorng problem that s dstnct from that of the classcal graph-colorng problem. Wth only three coarse-graned tme slots, VoIP capacty per AP can be boosted to 10 and 58 sessons n 80.11b and 80.11g, respectvely. Ths s another 35.3% mprovement over the three-frequency-channel clque-analytcal call admsson control strategy. The remander of the paper s organzed as follows. Secton explans how the 80.11 CSMA protocol would affect the VoIP capacty. In partcular, we show that the VoIP capacty can be eroded further n a sgnfcant way n the mult-cell settng. Secton 3 presents a call admsson strategy based on clque analyss of a graph-theoretc formulaton to confne nter-cell nterferences. Secton 4 consders usng a tme-dmenson approach n conjuncton wth the basc 80.11 CSMA to further mprove the VoIP capacty. Secton 5 concludes the paper. LOW VOIP CAPACITY OVER MULTIPLE WLANS.1 80.11 Protocols and VoIP VoIP packets are streams of packets contanng encoded voce sgnals. There are dfferent codecs for encodng voce sgnals. Take GSM 6.10 as an example. The voce payload s 33-byte and 50 packets are generated n each second. After addng the 40-byte IP/UDP/RTP header, the mnmum channel capacty to support a GSM 6.10 voce stream n one drecton (ether uplnk or downlnk) s 9.Kbps. An 80.11b WLAN n theory can support nearly 00 VoIP sessons (dvde 11Mbps by two tmes of 9.Kbps); and for 80.11g, more than 900 sessons (dvde 54 Mbps by two tmes of 9.Kbps). However, pror nvestgaton has shown that the actual VoIP capacty s severely lmted due to varous nherent header and protocol overheads. Wth the GSM 6.10 codec, for example, only 1 VoIP sessons can be supported n an 80.11b WLAN; and 60 sessons n 80.11g [1, ]. Pror work has prmarly focused on VoIP over an solated WLAN. When several WLANs are n the proxmty of each other, ther VoIP sessons may compete wth each other for artme usage f these WLANs use the same frequency channels. In 80.11b/g, for example, there are only three orthogonal channels, but t s not uncommon to see more than three overlappng WLANs n a buldng these days. We show n ths paper that the already-low effcency of VoIP over a sngle WLAN wll be further eroded n such a stuaton, so much so that only an average of less than two VoIP sessons per AP can be supported n a 5-by-5, 5-cell, 80.11b WLANs; and around 10 sessons n 80.11g.. Low VoIP Capacty over Multple WLANs Let us assume the allowance of packet-loss rate of 1% to 3%. Then, the mnmum channel capacty requrement s 8.3Kbps for GSM 6.10 codec. If both the uplnk and downlnk of a VoIP sesson can have throughputs exceedng ths benchmark, we say that the VoIP sesson can be supported n the WLAN. To evaluate the VoIP capacty over multple WLANs, we model a WLAN cell wth a regular hexagonal area of 50m on each sde. An AP s placed at the center of the cell. Any wreless clent staton nsde the cell wll be assocated wth the AP. The longest lnk dstance (d max ) s therefore 50m, whch s the data transmsson range (TXRange) for 80.11b assumed n NS. The transmsson ranges of APs n WLANs partally overlaps. Ths s also the case n practce. The crcles n Fg. 1 represent the coverage of certan WLANs (transmsson ranges of APs). In ths paper, nstead of usng crcles to model WLAN cell, we use hexagons for clearer llustraton. The results we derved from ths hexagonal-area model can be drectly appled to the crcular-area model. By placng the cells sde by sde, we form a D-by-D mult-cell topology, where D s the number of cells on each sde. Fg. 1 shows a 3-by-3 mult-cell topology. We consder the use of the basc mode [11] of 80.11 n ths paper because the short VoIP payload does not warrant the use of RTS/CTS. We assume the carrer sensng range (CSRange) of all wreless statons s 550m, the default n NS. A pont to note s that n real equpment, the operatng Fg. 1. A 3-by-3 mult-cell topology. magnfed vew d max d max AP

A. CHAN AND S. C. LIEW: VOIP OVER MULTIPLE IEEE 80.11 WIRELESS LANS 3 parameters are based on power thresholds rather than dstances. For example, for carrer sensng and for detecton, t s the power receved that matters rather than the dstance, and there s a one-to-one correspondence between dstance and power only f the wreless propagaton medum s spatally homogeneous. The mplementatons of the schemes (based on a graph-theoretc formulaton) expounded n ths paper are compatble wth the power-threshold nterpretaton and that the assumpton of a homogeneous medum s not necessary. Nevertheless, for convenence we wll contnue to use the dstances (e.g. TXRange and CSRange) to descrbe the system operaton. Thus, dstance s to be nterpreted n a vrtual sense, explaned as follows. Two wreless statons and j are sad to be separated by a vrtual dstance d, j f the power they receve from each other s P, j= Pj, = k / d α, j, where k s a constant, α s a reference (not the actual) path-loss exponent, assumng all statons use the same transmsson power. Applyng common k and reference to the whole network, we can then derve the vrtual dstance d, j from the measured power transferred, P, j. Gven a power threshold requrement, there s then a correspondng vrtual-dstance requrement. When we say the CSRange s set to d CS, we mean the carrer-sensng threshold power s set to P CS = k d α CS where α s the nomnal constant adopted (e.g., α = 4 ). In the subsequent dscussons, we assume there s an underlyng scheme to fnd out P, j, hence, d, j, so that we can assgn system resources (e.g., AP assocaton and tme-slot assgnment) accordng to d, j. To lmt our scope, however, we wll not dscuss the d, j dscovery algorthms here. The reader s referred to [1] for possble schemes. In short, the logcal correctness of the mplementaton of our schemes n ths paper does not depend on spatal homogenety. The performance, however, does depend on the α beng assumed. We ran smulaton experments on the D-by-D multcell topology for D =1,, 3, 4, and 5. In each run, wreless statons (VoIP sessons) are added one by one randomly assumng unform dstrbuton. When a partcular cell has 1 statons n the 80.11b case (60 statons n the 80.11g case), no more statons wll be added to that cell to ensure the capacty of the sngle solated WLAN case s not exceeded. Wth each addtonal VoIP sesson, NS smulaton s run and the throughput of each lnk recorded. When the next newly added VoIP sesson causes volaton of the packet-loss rate requrement (1% to 3%) by at least one of the sessons, we say that the capacty lmt has been exceeded. Ths corresponds to a smplstc call admsson scheme n whch upon unacceptable performance caused by the newly added sesson, the newly added sesson wll be dropped, and no more future sessons wll be accepted. We wll later consder a cleverer call admsson scheme based on a clque analyss of a conflct graph so that we can predct the performance before decdng whether to admt a call to get rd of the dsruptveness of addng a sesson only to drop t later. Table 1 summarzes our smulaton results. In Table 1, C D D s the total number of VoIP sessons that can be supported n a D-by-D mult-cell topology. Obvously, as D ncreases, more VoIP sessons can be supported. We further calculate C AP_D, the per-ap capacty n a D-by-D mult-cell WLAN, defned as follows: CAP _ D = CD D / D (1) We plot C AP_D aganst number of cells, D, n Fg.. We fnd that as the number of cells ncreases, per-ap capacty decreases. When the number of cells s 5,.e. D = 5, only 1.63 VoIP sessons can be supported by each AP n 80.11b. Compared wth the sngle-cell scenaro, where each AP can support 1 VoIP sessons n 80.11b, ths s a rather large penalty! Smlar capacty penalty s also observed n 80.11g, where only 10.34 sessons per AP can be supported when D = 5, as opposed to 55 n the snglesolated WLAN case when D = 1. 1 We also note n passng that for a network larger than the 5-by-5 network, most cells wll be surrounded by sx adjacent cells and there wll be proportonately fewer cells at the boundary, where they enjoy less nterference TABLE 1 VOIP CAPACITY OVER D-BY-D MULTI-CELL TOPOLOGIES D 1 3 4 5 Avg. C DxD n 11b 1.0 1.3 0.0 30.0 40.8 Avg. C DxD n 11g 55.0 79.6 131.8 07.6 58.4 Fg.. Per-AP capacty of D-by-D mult-cell WLAN. 1 The VoIP capacty of a sngle-cell 80.11g WLAN measured n smulaton s 55, not 60 as calculated n [1, ]. Ths s due to the smaller mnmum contenton wndow sze n 80.11g. But for consstency of comparson, we stll take 60 VoIP sessons as the theoretcal upper bound for VoIP capacty over sngle solated 80.11g WLAN n our followng dscussons

4 SUBMITTED TO IEEE TRANSACTIONS ON MOBILE COMPUTING from other cells. One can therefore expect the VoIP capacty per AP to drop even further when the dmensons are larger than 5-by-5. Indeed, n a 10-by-10, 100-cell, 80.11b WLAN, our smulaton shows that the per-ap capacty s only 1. sessons. STA AP AP1 STA1 d max d max Fg. 3. 3-channel assgnment n mult-cell WLAN..65d max Fg. 4. 7-channel assgnment n mult-cell WLAN..3 Applyng Frequency-Channel Assgnment To reduce the mutual nterference of neghborng cells, a quck soluton s to assgn dfferent frequency channels to dfferent cells, as n cellular telephone networks. If there are enough frequency channels, the neghbor cells that could nterfere wth each other through packet collsons and carrer sensng could be assgned dfferent frequency channels. Ths bols down to the same stuaton as n the sngle-cell case, so that the per-ap VoIP capacty n the mult-cell case s the same as that n the sngle-cell case. IEEE 80.11b/g has only three orthogonal frequency channels, and ths s not suffcent to completely solate co-channel nterference between cells. Fg. 3 shows that f we have only three frequency channels, then the nearest dstance between two cells usng the same channel (e.g. the two un-shaded hexagons n Fg. 3) s the same as the maxmum lnk length wthn a cell, d max. Snce these two cells may nterfere wth each other, the carrer-sensng range (CSRange) should be larger than 3d max to avod hdden-node [13] collsons between the two cells. To see ths, consder lnk (AP1, STA1) and lnk (AP, STA) of the two cells n Fg. 3, where the dstance between STA1 and STA s d max. The transmsson of STA1 wll nterfere wth the recepton of STA, and vce versa, so the two APs should be able to carrer sense each other for not startng transmsson on ther own lnks smultaneously. Hence, CSRange of 3d max s needed to avod hdden-node collsons between these two lnks. Snce a cell must now share artme wth other cells, the VoIP capacty per AP cannot be the same as that n the sngle-cell case. 80.11a, on the other hand, has twelve orthogonal channels. Fg. 4 shows that a 7-channel assgnment s suffcent for complete solaton of co-channel nterference. The nearest dstance between two cells usng the same channel s.65d max whch s larger than the mnmum CSRange, d max, used to prevent collsons wthn a cell. Thus, wth 7-channel assgnment, co-channel nterference can be completely solated. However, f we smply use 7 overlay networks n each cell (put 7 APs nsde each cell and operate n dfferent channels), the number of VoIP sessons that can be supported n each cell can be ncreased by 7 tmes of that n the sngle-channel mult-cell topology. Therefore, we fnd that the channel assgnment n Fg. 4 actually may not mprove the VoIP capacty on a per-frequency channel bass, although on a per-ap bass, t does. Furthermore, IEEE 80.11a operates n regulated frequency band (lcense s needed), so t s not commonly deployed. In short, the case of 80.11b/g n whch there are not suffcent frequency channels for complete elmnaton of co-channel nterference among adjacent cells wll be of much practcal nterest. Secton 3 consders how to perform call admsson wthn each frequency plane n whch cells assgned wth the same frequency channel may mutually nterfere wth each other. 3 CLIQUE ANALYSIS AND CALL ADMISSION To understand the cause for the heavy performance penalty n the mult-cell scenaro, we consder here a clque analyss based on a graph model that captures the conflct and nterference among the nodes. The clque analyss also suggests a call admsson methodology. Wth ths call admsson scheme, the VoIP capacty can be ncreased to.48 sessons per AP from 1.63 n the case of 5-by-5, 5-cell 80.11b WLAN. Ths consttutes a 5.1% mprovement. In the 5-by-5, 5-cell 80.11g WLAN, ths call admsson scheme ncreases the per-ap capacty by 36.75% from 10.34 to 14.14 sessons. 3.1 Conflct Graphs and Clques In our conflct graph model, vertces represent VoIP sessons (wreless lnks). An edge between two vertces means that the two VoIP sessons compete for the artme. In other words, they cannot transmt at the same tme. There are two reasons why they cannot transmt at the same tme. () Frst, nodes of the two sessons that are wthn the CSRange of each other wll be prevented by the 80.11 protocol from transmttng together. () Even f the two sessons are not wthn each other s CSRange, there may be mutual nterference between them so that ether one or both of ther transmssons wll fal f they transmt together: ths s due to the well-known hdden-node problem [13]. In ether case () or (), an edge s drawn between the two vertces of the VoIP sessons to mean that ther transmssons cannot use the same artme. A clque s a subset of vertces n whch there s an edge between any of two vertces [14]. The vertces n a clque compete for common artme. That s, the sum of the fractons of artmes used by the vertces should not exceed one.

A. CHAN AND S. C. LIEW: VOIP OVER MULTIPLE IEEE 80.11 WIRELESS LANS 5 In the sngle-cell scenaro, all clent statons are wthn the same cell and assocated wth the same AP. So, edges should be drawn among all vertces of the same cell. In an 80.11b sngle-cell WLAN, the maxmum clque sze s 1, because 1 VoIP sessons wll fll up all artmes [1, ]. In an 80.11g sngle-cell WLAN, the maxmum clque sze s 60. In a mult-cell WLAN, nstead of one clque, multple clques can be formed. To see ths, we need to nspect cases () and () mentoned above carefully. For case (), CSRange s usually a fxed value (assumng no power control). By default, t s 550m n NS. For case (), we consder the Interference Range (IR) defned as follows: IRk = (1 + ) d () k where IR k s the Interference Range of a node, k, (t can be a clent staton or an AP), d k s the length of the lnk assocated wth the node k; and s a dstance margn for nterference-free recepton wth typcal value of 0.78 [10, 13]. Wthn a radus of IR k, any other transmsson wll nterfere wth the node k s recepton of the sgnal. The maxmum lnk length wthn a cell s d max = 50m (shown n Fg. 1). The correspondng maxmum IR s therefore IRmax = 1.78 50m = 445m. In Fg. 5, the dstance between any two ponts of two dfferent shaded areas s at least 866m, whch s larger than both CSRange and IR. Therefore, there s no edge max between vertces of dfferent shaded cells, and the vertces of dfferent shaded cells belong to dfferent clques. From the dscusson above, we know that there s a maxmum clque sze (e.g., 1 and 60 for 80.11b and 80.11g sngle-cell WLANs respectvely) whch cannot be exceeded f satsfactory performance of VoIP sessons s to be attaned. To ncrease the VoIP capacty over multple WLANs, we need to pack the VoIP sessons (vertces) effcently. We dscuss ths call admsson control n the next subsecton. 3. Clque-Analytcal Call Admsson Control The ablty to predct whether a new VoIP sesson can be admtted wthout causng performance problem s mportant n call admsson. Prevous work [6, 15] showed that an addtonal VoIP sesson to a sngle-cell WLAN whch already reaches the maxmum capacty can degrade the performance of all other exstng VoIP sessons. Ths s also the case n mult-cell WLANs. In the 80.11b -by- network (n whch the maxmum capacty s around 1.3 as shown n Table 1), for example, f there are 14 VoIP sessons, three of them cannot fulfll the loss-rate requrement. That means only 11 sessons can be supported wth acceptable performance, as opposed to the 1 sessons that can be supported when there are exactly 1 sessons n WLAN. We consder a call admsson control mechansm based on the clque sze of the conflct graph. Let E be the set of vertces wth whch vertex has an edge. Let K be the set of all clques C to whch v, x v belongs, where x = 1,,.., K v s the ndex of the clques, and K s the total number of clques n K. Any par of clques n K must satsfy (3) below so that all clques n K are maxmal and not contaned n another clque [16]. As defned n (4), m s the sze of the largest clque n K. Cv, x C, v, y x y, 1 x, y K v (3) m v = max C (4) x v, x Fg. 6 gves an example of a conflct graph, where vertex v has the followng parameters. 1 E = { v, v, v, v } v1 3 4 5 Kv = { v 1 1, v, v3, v5},{ v1, v3, v4}, K v = 1.e., Cv 1,1 = { v1, v, v3, v5 }, Cv 1, = { v1, v3, v4} m v = 4 1 The pseudo code of the admsson control algorthm s gven n Algorthm I. There are three procedures n the algorthm: Procedures A, B and C. When there s a new call request (.e., a new vertex (VoIP sesson) wants to be added), Procedure A s frst executed, wheren a copy of the state ( Kv, m ) j v s frst saved n case the admsson of v fals and we need to revert to the orgnal j j state later. After that, Procedure B s executed. Procedure B updates the state ( Kv, m ) assumng the addton of j v v j. To satsfy (3), a functon call, NO_REDUNDANCY ( K v j ) V 4 866m V 1 V 3 866m V 5 V Fg. 5.. Multple clques may be formed n a 4-by-4 mult-cell WLAN Fg. 6. An example of a conflct graph.

6 SUBMITTED TO IEEE TRANSACTIONS ON MOBILE COMPUTING Algorthm I Procedure A Keep a copy of the state ( Kv, m ) for all exstng v ; j vj j Perform Procedure B; Procedure B for each vj E { for each Cv j, x K { v j f Cvj, x Ev then add v to C ; vj, x else add { Cv,, } j x Ev v to K ; v j } K = NO_REDUNDANCY v j ( K v j ); update m ; v j f mv j > C, max then reject v ; revert the state usng the copy stored n Procedure A; break out of Procedure B, and the algorthm s termnated; } // All vj E have been looked at f the algorthm comes to ths pont Perform Procedure C; Procedure C s admtted; K v = ; for each vj E { f v Cvj, x then C, x= C, vj, x Kv = K v C ; v, x } K = NO_REDUNDANCY ( K ); compute m ; s made. Algorthm II gves the pseudo code of NO_REDUNDANCY ( K v k ). Durng the updatng, Procedure B contnually checks to see f a pre-determned maxmum clque sze C max s exceeded so as not to volate the loss-rate requrement. If so, the algorthm s termnated and s rejected; n whch case the state saved n Procedure A s restored. If Procedure B successfully runs to the end wthout the C max beng exceeded, Procedure C s executed. Procedure C admts the new vertex v and calculates ( Kv, m ). v We have performed an experment n MATLAB to measure the executon tme of the call admsson control algorthm. The experment assumes the 5-by-5 mult-cell topology settng n Secton.. For smplcty, we frst consder the sngle-frequency channel case n whch all cells are assgned the same frequency. The next subsecton wll deal wth the mult-frequency case. Algorthm II NO_REDUNDANCY ( K v k ){ for each par Cv,, k x Kv C k vk, y K, x y { vk f Cv k, x Cv k, y then delete C ; vk, x } return ( K v k ); } We appled the algorthm to dentfy whch VoIP sessons out of a total of 300 (1 x 5, and 80.11b s assumed here) randomly placed (wth unform dstrbuton) potental sessons could be admtted n a 5-by-5 mult-cell WLAN wth the C max clque-sze constrant. Unlke the smplstc call admsson scheme earler, here when a sesson s rejected, the call admsson scheme contnues to consder a next sesson f the 300 sessons have not been exhausted. We ran several sets of experments for dfferent random node dstrbutons n a 5-by-5 WLAN. We used an ordnary personal computer wth 3.GHz CPU and 1G RAM to perform the experment. The results for C max = 8 and 1 are shown n Table. The total runtme s the total tme needed for the algorthm to go through all the 300 VoIP sessons. The average runtme s the tme needed to admt or reject a call (Total runtme / 300). We see that although the general clque problem s NPcomplete, the executon tme of our algorthm on the conflct graph that models 80.11 networks s acceptable. Based on the call admsson results n MATLAB, we then used NS to verfy whether the admtted calls meet the maxmum 3% packet-loss rate requrement n the smulaton. From Table, we fnd that when C max s 8, the average number of VoIP sessons admtted by our admsson control algorthm s 6.0. Ths s the average number of fve runs of the MATLAB experments. We ncorporated the correspondng sets of admtted VoIP sessons n fve dfferent runs of NS smulaton. In each run, all the admtted VoIP sessons can meet our packet-loss rate requrement. However, f nstead of settng C max to 8, we set t to 1, then nearly one thrd of VoIP sessons cannot meet the loss requrement. It s nterestng that for the large-scale mult-cell case, the maxmum clque sze that should be mposed s 8 rather than 1 (recall that 1 s the maxmum clque sze n the sngle-cell topology f 80.11b s assumed) f loss-rate requrement s to be satsfed. Ths s TABLE EXPERIMENT RESULTS OF APPLYING THE ADMISSION CONTROL ALGORITHM ON 5-BY-5 MULTI-CELL WLAN C max Total Runtme (s) Average No. of Sessons Admtted Average Runtme (s) 8 48.6 6.0 0.160 1 11.1 70.4 0.374

A. CHAN AND S. C. LIEW: VOIP OVER MULTIPLE IEEE 80.11 WIRELESS LANS 7 perhaps due to the nteracton and couplng among dfferent clques caused by the 80.11 MAC protocol. In other words, 80.11 MAC may not acheve perfect schedulng n whch the artme usage wthn each clque s 100% tghtly packed. Ths motvates us to explore the use of tme-slot schedulng over the basc CSMA 80.11 MAC n Secton 4 for performance mprovement. Wth C max = 8, the 5-by-5 mult-cell WLAN can support.48 sessons (6.0/5) per AP, yeldng a 5.1% mprovement over the smplstc call admsson scheme n Secton. The clque-analytcal call admsson control works smlarly well n 80.11g WLANs. For the 5-by-5 mult-cell 80.11g WLAN, wth C max = 44, the per-ap capacty can be ncreased from 10.34 to 14.14 VoIP sessons, yeldng a 36.75% mprovement. 3.3 Clque-Analytcal Admsson Control n Three- Frequency-Channel WLANs In ths subsecton, we explore the mpact of multple frequency channels on VoIP capacty over multple WLANs. The avalablty of multple frequency channels allows us to separate the cells usng the same frequency by a longer dstance. Farther separaton of cells that use the same frequency leads to less conflct among transmssons of dfferent VoIP sessons (although not elmnatng conflcts entrely). Consequently, fewer edges are formed n the correspondng conflct graph. Consder the mult-cell topology n Fg. 7, where we apply the three frequency-channel assgnment (as n Fg. 3). In Fg. 7, the shaded cells use the same frequency channel, whle the un-shaded cells use the other two frequency channels. Although the sze of the topology n Fg. 7 s comparable to the 5-by-5, 5-cell WLAN we descrbed n prevous sectons, the three frequency channels help to reduce conflcts and ncrease the number of VoIP sessons that can be supported by each AP. NS smulatons show that applyng the clque-analytcal admsson control to the three frequency-channel layout can boost the per-ap capacty to 7.39 VoIP sessons n 80.11b; and 44.91 VoIP sessons n 80.11g. These are respectvely.98 and 3.17 tmes of the per-ap capacty n 80.11b and 80.11g 5-by- 5, 5-cell WLAN where sngle-frequency channel s used. In the next secton, we explore tme dvson on 80.11 MAC whch can elmnate all HN and allevate the exposed-node (EN) problem. The VoIP capacty over multple WLAN can be further mproved. Fg. 7. 3-frequency-channel assgnment s appled on multple WLANs. 4 TIME-DIVISION CSMA MAC So far we have learned that there s a sgnfcant capacty penalty when we move from the sngle-cell WLAN scenaro to the mult-cell WLAN scenaro. We have also learned that the extremely low VoIP capacty over multple WLANs s due to the mutual nterference from neghborng cells, and such mutual nterference cannot be completely solated even wth careful frequency channel assgnment. Although IEEE 80.11e has been standardzed recently to support QoS n WLAN [3], t only focuses on the sngle-cell stuaton. In ths secton, we explore addng the tme-dvson approach to the basc 80.11 CSMA protocol. We show that the ntegrated tme-dvson-csma approach can potentally ncrease the VoIP capacty over multple WLANs qute sgnfcantly. Selected prevous work has consdered Tme Dvson Multple Access (TDMA) MAC. However, ther focus s on the sngle WLAN case [17-0]. In addton, CSMA s proposed to be replaced entrely by TDMA [1-3].e., the motvaton was not to explore a soluton workable wthn the context of the wdely-deployed 80.11 technology. For VoIP applcatons wthn 80.11, each transmsson conssts of a very small packet (relatve to the raw data rate and the varous packet headers), and fne TDMA requres tght synchronzaton among the statons, whch can n turn cause throughput degradaton. In ths secton, our prmary focus s on the prncple of coarse tme dvson, n whch a relatvely large tme slot s allocated to a group of statons. The statons of the same tme slot then contend for channel access usng the orgnal 80.11 CSMA scheme. We () lay out and nvestgate a graphtheoretc problem formulaton that captures the prncple of ntegratng coarse tme dvson wth CSMA n Subsecton 4.1; and () provde a feasblty nvestgaton of the approach wthn the context of 80.11 n Subsecton 4.. 4.1 Coarse-grand Tme-Dvson Mulple Access In Secton.3, we have dscussed 3-frequency-channel assgnment n 80.11b/g WLAN (see Fg. 3) and argued that frequency-channel assgnment alone cannot completely solate co-channel nterference. Secton 3.3 apples the clque-analytcal call admsson control to boost per- AP VoIP capacty n 3-frequency-channel WLAN; however, the co-channel nterference from dfferent cells s stll not solated entrely. The goal of Coarse-graned Tme-Dvson Multple Access (CoTDMA) s to remove such co-channel nterference. In partcular, we mpose a restrcton such that two statons n dfferent cells that nterfere wth each other would be allocated dfferent tme slots or dfferent frequency channels. 4.1.1 Basc Ideas of CoTDMA We frst explan the concept usng the smplfed scenaro depcted n Fg. 8. In Fg. 8, n addton to the 3-frequencychannel assgnment, we dvde each cell nto sx sectors and assgn a dstnct tmeslot to each sector. The shaded cells use the same frequency channel, and the statons

8 SUBMITTED TO IEEE TRANSACTIONS ON MOBILE COMPUTING 3 1 4 5 6 3 1 4 5 6 3 1 4 5 6 3 1 4 5 6 3 1 4 5 6 3 1 4 5 6 3 1 4 5 6 Fg. 8. 6-tmeslot assgnment n addton to 3-frequency-channel assgnment n multple WLANs. wthn each sector use the same tme slot; the number n each sector ndexes the tme slot assgned to that partcular sector. The frequency and tme-slot assgnments n Fg. 8 are such that dfferent sectors do not nterfere wth each other because they are ether actve n dfferent tme slots, n cells of dfferent frequences, or suffcently far apart from each other. Ths allows us to shorten the CSRange to d max, the dstance from the AP located at the center of each cell to the farthest corner n the cell. The nearest staton wth the same frequency and tme-slot assgnment n neghborng cells s at least d max away, whch s larger than both the CSRange and IR defned n (). Therefore, the co-channel nterference from neghborng cells s completely solated. Note that wthn each sector (tme slot), there may stll be multple statons, and the orgnal 80.11 CSMA scheme s used to coordnate transmsson among these statons. In the best-case scenaro, the VoIP capacty per AP n multple WLANs can be the same as that n the sngle solated WLAN case. To see ths, consder 80.11b: f each sector has two statons, then we can have a total of 1 VoIP sessons per AP. Although the sectorzed cells n Fg. 8 llustrate the concept of CoTDMA, t has two mplementaton dffcultes: 1) tme-slot assgnment requres the knowledge of the locatons of the ndvdual statons; ) f the statons are not evenly dstrbuted across the sectors, then t wll not be as effectve as the best-case scenaro mentoned above. In the followng, we present a graph model of CoTDMA to solve these problems. The graph model presented also gves a framework for performance analyss of CoTDMA. For convenence, we wll contnue to descrbe the system operaton n terms of dstances such as CSRange and IR. As stated n Secton., mplementaton based strctly on the geometrc dstance nterpretaton s unnecessary once we move on to the graph-theoretc formulaton here. A mappng of dstances to power thresholds s all that s needed. Defnton 1: In CoTDMA, m frequency channels and n tme slots are assgned to the VoIP sessons. In each cell, at most k VoIP sessons are actve n each tme slot, where k = C / AP _1 n, and C AP_1 s the average per-ap capacty n a sngle solated WLAN. 4.1. Conflct Graph Modelng of CoTDMA Accordng to Defnton 1, the parameter m s the number of orthogonal frequency channels avalable. For example, m = 3 n 80.11b and 80.11g. If n = 6, then k = n 80.11b, and k = 10 n 80.11g, snce the respectve C AP_1 are 1 and 60. We wll look at the system performance as a functon of n shortly. Frst, we formulate the correspondng graphtheoretc colorng problem. Colorng s a well-known problem n graph theory. However, the assgnment problem n CoTDMA does not drectly map to the classcal colorng problem n graph theory. For CoTDMA, a modfed constructon of the conflct graph, as well as a modfed colorng problem, s needed to reflect the specfcs of 80.11 CSMA scheme, as detaled below. Instead of pre-assgnng the three frequency channels as n Fg. 8, let us frst set up a general framework whch ntegrates the frequency channel assgnment and tmeslot assgnment. In the CoTDMA conflct-graph model, we use two layers of colorng. The frst-layer colors represent frequency channels and the second-layer colors represent tme slots. The frst-layer colorng s appled at the cell level (assumng all nodes wthn the cell use the same frequency channel n a statc manner) whle the second layer colorng s appled at the staton (vertex) level. In the followng, we frst state the constrants of our colorng problem under the context of 80.11, and then descrbe how to capture the constrants n the conflct graph. Vertces are assocated wth the clent statons n the followng: Constrant 1: The number of avalable frst-layer colors s m, and the number of avalable second-layer colors s n (see Defnton 1) Constrant : All vertces assocated wth the same AP (wthn the same cell) must have the same frst-layer color. Constrant 3: Wthn a cell, there can be at most k vertces assgned wth the same second-layer color. Furthermore, the vertces assgned the same second-layer color n a cell must be wthn the CSRange of each other. For vertces wthn the same cell, t s obvous that the assocated statons cannot transmt together snce one end of the lnks s always the AP. The ssue for vertces wthn the same cell s that whether the CSRange of vertces (clents) can cover each other, hence Constrant 3.e., f two vertces are not wthn each other s CSRange, then carrer sensng between them does not work, and therefore they should be assgned dfferent tme slots to avod the hdden-node phenomenon. Constrant 4: Consder two vertces of dfferent cells, v and v. They conflct wth each other and must be as- j

A. CHAN AND S. C. LIEW: VOIP OVER MULTIPLE IEEE 80.11 WIRELESS LANS 9 sgned dfferent (frst-layer color, second-layer color) combnatons f one or more of the followng nequaltes below are satsfed: CSRange mn( d, d, d, d ) (5) v, vj v, vj' v', vj v', vj' IRv > mn( d,,, ') v v d j v vj IRv ' > mn( d ',, ', ') v v d j v vj IRv > mn( d,, ', ) j v v d j v vj IR > mn( d, d ) vj' v, vj' v', vj ' where and v j are the correspondng APs that and v j assocated respectvely. Note that both (5) and (6) descrbe the condtons under whch smultaneous transmssons are not possble. However, there s a subtlety. Inequaltes (6) capture the condtons that wll lead to collsons. Inequalty (5), on the other hand, only says that the CSMA mechansm wll prevent the statons from transmttng together that s, strctly speakng, f (5) s satsfed, the statons could stll be assgned the same color combnaton, and the CSMA mechansm smply prevents smultaneous transmssons (f we dd that, Constrant 3 would need to be modfed to encompass the overall network). Constrant 4, however, dsallows that as a desgn choce to smplfy thngs. The reasons are as follows: () If dfferent color combnatons are used whenever (5) s satsfed for two vertces n dfferent cells, then CSMA n dfferent cells wll then be decoupled n each of the tme slots under CoTDMA, obvatng the need for nter-cell CSMA. () When nequaltes (5) and (6) are mposed on CoTDMA colorng, we may decrease CSRange to only meet the need of ntra-cell CSMA, and there s no need for CSRange to be large enough to ensure proper CSMA operaton across cells. Ths has the advantage of reducng exposed nodes (EN) across cells [1], hence ncreasng spatal re-use. Later n ths secton, we wll explore the optmal value for CSRange through smulatons. Capturng Constrants to 4 n Conflct Graph To capture Constrant, we could assgn an AP_ID to the vertces n accordance wth the APs to whch they assocate. Vertces wth the same AP_ID must be gven the same frst-layer color. For Constrants 3 and 4, an edge between two vertces means they must be assgned dfferent (frst-layer color, second- layer color) combnaton. To capture constrant 3, there s an edge between two vertces v and v of the same cell f j d v, vj (6) > CSRange (7) Two vertces that are wthn the CSRange of each other could be assgned the same or dfferent second-layer color. However, there can be at most k vertces wth the same second-layer color wthn a cell. Constrant 4 can be captured by drawng an edge between two vertces of dfferent AP_ID f there s a conflct relatonshp under nequaltes (5) and (6). To avod confuson, t s worth emphaszng agan that between vertces of dfferent AP_ID (from dfferent cells), we use (5) and (6) for the edge formaton crtera. For vertces of the same cell, we use (7) for the edge formaton crtera. A formulaton of the CoTDMA problem s as follows: -Color Assgnment Problem: Assgn (frst-layer color, second-layer color) to the vertces subject to constrants 1 to 4 such that the total number of vertces successfully colored s maxmzed. Fg.9 llustrates the dea of CoTDMA. APs (trangles) are at the center of the cells, clent statons (crcles) n the same cell are assocated wth the same AP. The sold-lne cells use the same frequency channel whle the dottedlne cells use the other two frequency channels. For smplcty, we assume the standard 3-frequency-channel assgnment here for the frequency channel assgnment n CoTDMA. It s worth notng that ths standard 3- frequency-channel assgnment s not a must for CoTDMA. In ths case, we only draw a conflct graph for secondlayer colorng. Due to the standard 3-frequency-channel assgnment, the vertces represent clents N1 to N6 all have the same frst-layer color. Here we assume 80.11b and n = 6 (whch mples k = gven a capacty of 1 VoIP sessons per cell), and CSRange = d max. Fg. 10 shows the correspondng conflct graph together N1 N6 N4 N5 N d max N3 d max Fg. 9. A topology of VoIP over multple WLANs wth 3-channel assgnment. (1, 1) v 1 v 6 (1, 3) (1, 1) v 4 v 5 v (1, ) (1, ) v 3 (1, 1) Fg. 10. A conflct graph wth frst-layer and second-layer colors for the network n Fg. 9.

10 SUBMITTED TO IEEE TRANSACTIONS ON MOBILE COMPUTING wth the colorng. The bold numbers n parentheses besde the vertces are the frst-layer colors whle the other numbers are the second-layer colors. Accordng to Constrant 3, for second-layer colorng, N1 and N4 are assgned COLOR1, N and N5 are assgned COLOR, and N6 are assgned COLOR3. N3 and N5 may nterfere wth each other under (5) and (6), so an edge s drawn between them. Accordng to Constrant 4, the two vertces must be assgned wth dfferent color (frst-layer color, secondlayer color): n Fg. 10, COLOR1 s assgned to N3 and COLOR s assgned to N5. 4.1.3 Parameter Values n CoTDMA An mportant parameter n CoTDMA s n, the number of the tme slots (second-layer colors) avalable for the system. From Defnton 1, k, the number of VoIP sessons that can be actve n the same tme slot n a cell s set accordngly to n. In the above example, we set n = 6 and k =. The value of n drectly affects the number of vertces that can be successfully colored. A larger n (.e. a smaller k) means more fnely-dvded tme slots. In the extreme case, k = 1 (n = C AP_1,.e. n = 1 for 80.11b and n = 60 for 11g), whch s a pure TDMA scheme. In ths case, every VoIP sesson n a cell s assgned a dstnct tme slot. Hence, no carrer sensng s requred for accessng the medum. From the graph-theoretc colorng vewpont, fne TDMA as such would allow us to ncrease the number of vertces successfully colored n the two-layer-color assgnment problem defned above. However, fne TDMA has an mplementaton cost not captured n the colorng problem namely, there s the need for a guard tme as we swtch from slot to slot. Ths mplementaton ssue wll be further dscussed n the next subsecton. For the tme beng, suffce t to say that we are nterested n makng k as large as possble (.e. makng n as small as possble) whle retanng the performance results of the case where k s set to 1. Another mportant parameter n CoTDMA s CSRange. As mentoned n the explanaton of Constrant 4 n Subsecton 4.1., CoTDMA allows us to decrease CSRange and the carrer sensng mechansm needs only to work properly wthn a cell. However, f the CSRange s too small, a clent may only carrer-sense few other clent statons wthn the cell. Accordng to constrant 3 (and nequalty (7)), small CSRange may cause more ntra-cell edges n the conflct graph that restrct colorng freedom. For llustraton, let us frst consder settng CSRange to be the sector dameter, as descrbed below. Each cell s dvded nto n sectors n a fashon that generalzes Fg. 8. CSRange could be set to cover the maxmum dstance between any two ponts wthn a sector. Let us refer to the maxmum dstance as the dameter of the sector. For example, f 80.11b s assumed, C AP_1 = 1. For n = 1, 6, 4, 3,, and 1, the correspondng CSRange could be set as shown n Table 3. In Table 3, we also show the correspondng values of k accordng to Defnton 1. If 80.11g s assumed, the same set of CSRange values should be used f n = 60, 6, 4, 3, and 1 (as n 80.11g, C AP_1 = 60). The sector-dameter s a worst-case settng n the sense that t assumes that there exst actually two statons at the extreme corners of a sector that defne the dameter, whch s rarely the case. It would therefore be of nterest to explore tunng the CSRange to maxmze the number of vertces can be successfully colored n the conflct graph. We present the smulaton results based on the sectordameter settngs as well as smaller-than-sectordameter settngs below. We frst show the results of sector-dameter settngs. Based on the parameters n Table 3, we have performed MATLAB experments assumng 80.11b (C AP_1 = 1) and 80.11g (C AP_1 = 60). For smplcty, the standard 3- frequency-channel assgnment s assumed. Snce the frstlayer colors are pre-fxed n ths experment, only secondlayer colorng (tme-slot assgnment) s consdered. We use hexagonal cells to model WLAN, and 1 (for 80.11b) or 60 (for 80.11g) wreless clent statons are randomly placed nsde each cell wth unform dstrbuton. We use a heurstc algorthm of Welsh and Powell [4] to color the conflct graph. We add our colorng constrants 1 to 4 to talor the algorthm to CoTMDA. The algorthm of Welsh and Powell does not gve optmal graph colorng n general, but the effect of n s already qute pronounced even wth the smple heurstc. Fg. 11 and 1 respectvely show the percentages of vertces that can be colored when C AP_1 = 1 and C AP_1 = 60. In the experment, CSRange changes accordngly to n as dctated by Table 3. For each n, we ran fve experments wth dfferent node dstrbutons. In the fgures, crcles are the average percentage of successfully colored vertces. In general, as n ncreases, more vertces can be successfully colored. Although not shown n the fgures, the two cases for n = 1 n 80.11b and n = 60 n 803.11g have 100% of ther vertces successfully colored n all runs of our experments. But even for n as small as 3 or 4, the performance of CoTDMA s already very good (100% of vertces are colored n 80.11b, whle more than 95% of vertces are colored n 80.11g). Wth smaller n, the overhead of guard-tme for swtchng between tme slots can be reduced (to be elaborated shortly) and n = 3 or 4 may offer the best desgn tradeoff. TABLE 3 VALUE OF n, k AND CORRESPONDING CSRange IN 80.11b n 1 3 4 6 1 k 1 6 4 3 1 CSRange 13 dmax d max 3d max 7 d max dmax 0

A. CHAN AND S. C. LIEW: VOIP OVER MULTIPLE IEEE 80.11 WIRELESS LANS 11 Fg. 11. Percentage of colored vertces when CSRange s set as sector-dameter and C AP_1 =1 (80.11b). Fg. 13. Average percentage of colored vertces as CSRange changes when C AP_1 =1 (80.11b). Fg. 1. Percentage of colored vertces when CSRange s set as sector-dameter and C AP_1 =60 (80.11g). Fg. 14. Average percentage of colored vertces as CSRange changes when C AP_1 =60 (80.11g). In the next set of experments, for each run, we set a fxed CSRange for all values of n. Across dfferent runs, we vary CSRange. Fg. 13 and 14 show the average percentage of successfully colored vertces as a functon of n under dfferent fxed CSRange. We fnd that when CSRange s d max, the overall system performance s poor because many edges are formed wthn a cell (accordng to (7)). As CSRange ncreases, the overall system performance mproves. But beyond certan pont (e.g., 1.5d max n the fgures), the overall system performance drops agan. Ths phenomenon reveals the tradeoff between ntra-cell optmalty and nter-cell optmalty. When CSRange s too large, say d max, many edges are formed between vertces of dfferent cells (accordng to (5)), leadng to an ncrease of exposed nodes. Through expermentaton, we fnd the optmal CSRange whch yelds the best system performance s around 1.637d max. Indeed, we could use ths settng for dfferent n values wth reasonably good results (see Table 4). Our experment results show that even though the freedom of the two dmensons of frequency and tme n TABLE 4 SYSTEM PERFORMANCE FOR CSRange = 1.637d max n 1 3 4 6 CAP_1 % vertces successfully colored n 80.11b % vertces successfully colored n 80.11g 80.0 98.6 100 100 100 100 69.8 93.0 99.6 100 100 100 CoTDMA has not been fully utlzed (due to our assumng the fxed 3-frequency-channel assgnment), CoTDMA can generally color large portons (over 90%) of the vertces of the conflct graph when n 3 for both 80.11b and 80.11g. The performance can be even better when approprate ( optmal ) CSRange s set. By contrast, wthout CoTDMA, and wth 3-frequency-channel assgnment alone (n = 1), we fnd that only around 60% of the total capacty can be utlzed.

1 SUBMITTED TO IEEE TRANSACTIONS ON MOBILE COMPUTING 4. Possble Realzatons of Tme Dvson Multple Access wthn Exstng IEEE 80.11 Standards Most prevous work [18-3] consdered propretary protocols for mplementng TDMA on wreless networks. Our focus here s to mplement tme-slot assgnment wthn the framework of the IEEE 80.11 standard. A most crtcal ssue s how to realze the concept of tme slot wthn the 80.11 CSMA structure. A possblty s to make use of the sleepng mode n 80.11, whch was orgnally desgned for power conservaton purposes. In the sleepng mode, beacon frames are used for synchronzaton. Accordngly, n CoTDMA, all statons could wake up around the beacon tme for synchronzaton. As llustrated n Fg. 15, n CoTDMA, wthn each beacon nterval (BI) between the end of a beacon and the begnnng of the next beacon, the tme s dvded to C frames (cycles), each of duraton T. Wthn each frame, there are n tme slots, each of duraton t. A VoIP stream transmts one VoIP packet n each frame. The offset from the end of the beacon to the begnnng of the th frame s ( 1) T. A staton that has been assgned tme slot j s to be awake wthn a BI only durng the tme ntervals specfed by [ ( 1) T + ( j 1) t, ( 1) T + j t ), = 1,,, C. Other than these tme ntervals and the beacon tme, the staton sleeps. Guard-Tme Overhead n CoTDMA The guard tme should be set to the duraton of one VoIP packet. That s, no packet transmsson should be ntated wthn the current tme slot when the begnnng of the next tme slot s only a guard tme away. Ths s to ensure packet transmsson wll not straddle across two tme slots. Let r be the maxmum number of VoIP packets (ncludng the CSMA overhead) that could be trans- t mtted and receved wthn each tme slot by all VoIP sessons whch are actve n that partcular tme slot. Then, RVoIP r = t CAP_1 10 nc (8) where R VoIP s the number of VoIP packets generated per second n a partcular VoIP codec. The factor of s due to each VoIP sesson havng a downstream and an upstream flow. The factor of 1/10s due to the default 0.1s separaton tme between two beacons. The guard tme overhead s a constant of one VoIP-packet duraton, so that the tme-slot effcency s ( r t 1)/ r t. Thus, smaller r t gves rse to lower effcency, whch n turn results n lower capacty. From (8), we can see that t s desrable to Beacon Frame B Δt ΔT=nΔt.. CΔT (BI).... Fg. 15. Frame structure of CoTDMA when t s mplemented usng 80.11 sleepng mode. B. set n and C as small as possble. From the smulaton experments n the prevous subsecton, however, we need to make sure that n 3 (see Subsecton 4.1) so as to make sure most sessons wthn the system capacty lmt can be admtted. Whle call admsson consderaton mposes a lmt on how small n can be, the delay budget consderaton mposes a lmt on how small C can be, as explaned n the next few paragraphs. In other words, the factors that bound on the sze of n and C are dfferent. In CoTDMA, the maxmum delay for a VoIP packet s T + B, where B s the duraton of the beacon (see Fg. 15). To see ths maxmum delay, let us consder the staton beng assgned tme slot j. In the worst case, t could generate a packet n the last frame wthn a BI just slghtly after tme slot j ends, thus mssng t. The earlest tme for the next tme slot j (n the next BI ) s T + B t after that. Wthn ths next tme slot j, n the worst case, the packet s sent just before the end of the tmeslot (due to the CSMA contenton wth other statons assgned the same tme slot). If we assume the packet wll be dscarded f t fals to be sent out by ths tme - so as to make way for a newly generated packet from the same VoIP sesson - the maxmum delay s then T + B. Thus, T + B DB (9) where DB s the delay budget. Snce C T = BI (see Fg. 15), we have BI BI / C + B DB => C (10) DB B Suppose we set a local delay budget of 30ms for VoIP applcatons [1]. A typcal value of B s 0.5ms. Wth the default separaton tme between two beacons of 100ms, a beacon nterval, BI, s 99.5ms. Hence, the number of frames, C, n a beacon nterval, s at least 99.5/9.5 = 3.37. C should be a postve nteger as well, so the smallest C s 4. Assumng n = 3, C = 4, and GSM 6.10 codec and 80.11b (R VoIP = 50, C AP_1 = 1), from (8) we fnd that 90% of capacty s utlzed. That s at most 1 90% = 10 VoIP sessons can be admtted per AP n 80.11b networks. If 80.11g s assumed (C AP_1 = 60), 98% of capacty s utlzed. That s 58 VoIP sessons can be admtted per AP n 80.11g networks. Take 80.11b WLAN. Wth 3-frequency-channel CoT- DMA, the per-ap capacty over multple WLANs s 10 VoIP sessons. Ths s a 35.3% mprovement over the per- AP capacty of 7.39 sessons for the three-frequencychannel clque-analytcal call admsson control strategy n Secton 3.3. Another possblty for mplementng the concept of tme-dvson s to use the pollng mechansm of Pont Coordnaton Functon (PCF) [5] to mtate the tme slot assgnment n CoTDMA. In PCF, traffc s scheduled by the AP, so that no extra guard-tme overhead s needed for tme slot swtchng. If we assume PCF, we could set n = C AP_1 (.e., k = 1) n our CoTDMA call admsson scheme,

A. CHAN AND S. C. LIEW: VOIP OVER MULTIPLE IEEE 80.11 WIRELESS LANS 13 so that we essentally have the extreme case of Fnegraned Tme Dvson Multple Access (FTDMA). In PCF, an AP mantans a pollng lst contanng all the wreless statons n ts WLAN. In the contenton-free perod, the AP polls the statons on the pollng lst one by one, snce only one VoIP sesson s actve n each tmeslot n a cell n FTDMA. Only the polled staton can transmt packets to AP (the downstream packet s attached n the pollng packet). The poston of a staton n the pollng lst corresponds to the tme slot assgned to the staton. In ths way, statons whch may nterfere wth each other n adjacent cells wll not be polled at the same tme. Wth FTDMA, the number of vertces successfully colored can be ncreased. Furthermore, snce only one VoIP sesson s actve n each tmeslot n a cell, carrer sensng mechansm can be deactvated. Wthout backoff countdown n contenton perod, C AP_1, the average per-ap capacty n sngle solated WLAN, can be boosted n PCF (n 80.11b, C AP_1 ncreases from 1 to 17, whle n 80.11g, C AP_1 ncreases from 60 to over 90). A major concern, however, s that PCF s seldom used n practce and many 80.11 devces do not support t unlke DCF, the robustness of PFC n feld deployment has not been well tested. 5 CONCLUSION In ths paper, we have shown that when there are multple 80.11 WLANs wthn the vcnty of each other, the already low VoIP capacty n the sngle solated WLAN case (around 1 and 60 VoIP sessons per AP n 80.11b and 80.11g, respectvely) s further eroded n a very sgnfcant manner, so much so that less than 1% goodput can be supported. For example, n 80.11b, less than VoIP sessons per AP can be supported whle throughput computaton based on raw bandwdth of the WLAN yelds 00 sessons per AP. In 80.11g, around 10 VoIP sessons per AP can be supported whle 900 sessons can be yelded from throughput computaton. The dsmal performance and neffcency mply that there s much room for mprovement wthn the 80.11 standard as far as the support for VoIP s concerned. The low VoIP capacty n the sngle WLAN [1, ] s due largely to header overheads, and packet aggregaton [1, ] s an effectve soluton to reduce the header penalty. The further degradaton of VoIP capacty n the multple- WLAN case, however, s due to nterferences among the WLANs, and requres addtonal solutons. Ths paper suggests a two-pronged approach 1) call admsson control; and ) vrtual channelzaton. Regardng 1, we have formulated a clque-analytcal call admsson control algorthm, and shown that (compared wth a smplstc call admsson scheme) t can mprove the VoIP capacty n a 5-by-5, 5-cell 80.11b WLAN by 5.1% from 1.63 sessons to.48 sessons per AP. The mprovement s 36.8% n 11g. If three orthogonal frequency channels are used, such as are avalable n 80.11b/g, the capacty can be ncreased to 7.39 (11b) and 44.91 (11g) VoIP sessons per AP by careful frequencychannel assgnment to the cells. Regardng, the three orthogonal frequency channels n 80.11b/g are not enough to completely solate nterferences among WLANs. Ths n turn requres the CSRange of 80.11 to be set to be rather large to prevent packet collsons; but dong so also ncrease the exposednode problem that degrades the VoIP capacty. In ths paper, we have shown that vrtual channels (or tmeslot channels) could be created to combat ths problem. By assgnng the vrtual channels judcously to the VoIP statons, we could effectvely solate the nterferences between cells. Specfcally, we have proposed coarsegraned tme-dvson multple access (CoTDMA) as a means for vrtual channelzaton wthn the basc 80.11 CSMA protocol, so as to mantan compatblty wth the wdely-deployed 80.11 technology to a large extent. In CoTDMA, the tme dmenson s dvded nto multple coarse tme slots; multple VoIP sessons are assgned to each tme slot, and the VoIP sessons assgned to the same tme slot n a WLAN make use of the basc 80.11 CSMA protocol to coordnate channel access. The basc dea s that VoIP sessons of adjacent WLANs that may nterfere wth each other n a detrmental way should be assgned dfferent tme slots or frequency channels, so that VoIP sessons of dfferent cells do not need to use CSMA to coordnate transmssons among them, and that CSMA s n effectve use only wthn a cell. We note that CoTDMA s fundamentally a technque n whch statons contendng for a common resource (.e., artme) are compartmentalzed nto subsets so that only statons wthn each subset contend wth each other. The parttonng s done n such a way that the subnetwork consstng of statons wthn each subset s less susceptble to detrmental nterference/carrer-sensng pattern so that the resource could be used more effcently. Ths prncple s applcable not just to voce traffc, but to wreless networkng n general wth or wthout voce traffc. As further work, we note that the graph-theoretc formulaton of the two-layer colorng problem n Secton 4 s rather general, and could be explored for the general case. In the smulaton experments of ths paper, even though the freedom of the two dmensons of frequency and tme n CoTDMA have not been fully exploted (due to the assumpton of a standard 3-frequency channel assgnment to the WLANs), we fnd that CoTDMA could already mprove the VoIP capacty meanngfully. Our results ndcate that wth a small number (3 to 4) of coarse tme slots n CoTDMA, the per-ap VoIP capacty can be ncreased to 10 sessons n 80.11b and 58 sessons n 80.11g (another 35.3% and 9.15% mprovement over clque-analytcal admsson control wth three orthogonal frequency channels, respectvely). Last but not least, we note that although the effect of mutual nterference on VoIP over mult-hop mesh networks has also been consdered [7-9], our proposed CoTDMA scheme s targeted at solvng mutual nterference n nfrastructure networks, the predomnant archtecture deployed

14 SUBMITTED TO IEEE TRANSACTIONS ON MOBILE COMPUTING n the feld today. Havng sad that, t s stll nterestng to explore whether and how smlar approaches could be appled to mult-hop networks. REFERENCES [1] W. Wang, S. C. Lew, V. O. K. L, Solutons to performance problems n VoIP over 80.11 wreless LAN, IEEE Trans. on Vehcular Technology, vol. 54, no. 1, Jan. 005. [] W. Wang, S. C. Lew, Q. X. Pang, and V. O.K. L, A multplexmultcast scheme that mproves system capacty of Voce- over- IP on wreless LAN by 100%, The Nnth IEEE Symposum on Computers and Communcatons, Jun. 004. [3] IEEE Std. 80.11e, Part 11: Wreless LAN Medum Access Control (MAC) and Physcal Layer (PHY) specfcatons. Amendment 8: Medum Access Control (MAC) Qualty of Servce Enhancements, Nov. 005. [4] D. P. Hole, F. A. Tobag, Capacty of an IEEE 80.11b wreless LAN supportng VoIP, Proceedng of the IEEE ICC 04, vol. 1, pp. 196-01, Jun. 004. [5] F. Anjum et al, Voce performance n WLAN networks an expermental study, Proceedngs of IEEE Globecom 03, vol. 6, pp. 3504-3508, Dec. 003. [6] S. Garg, M. Kappes, An expermental study of throughput for UDP and VoIP traffc n IEEE 80.11b networks, Proceedngs of IEEE WCNC 03, vol. 3, pp. 1748-1753, Mar. 003. [7] H. Wu et al., SOFTMAC: Layer.5 collaboratve MAC for multmeda support n multhop wreless networks, IEEE Transacton on Moble Computng, vol. 6, no. 1, Jan. 007. [8] D. Nculescu et al., Performance of VoIP n a 80.11 wreless mesh network, Proceedngs of IEEE INFOCOM 006, Apr. 006. [9] S. Ganguly et al., Performance optmzatons for deployng VoIP servces n mesh networks, IEEE Journal on Selected Areas n Communcatons, vol. 4, no. 11, Nov. 006. [10] The Network Smulator ns, http://www.s.edu/ nsnam/ns/ [11] M. S. Gast, 80.11 Wreless Networks: Defntve Gude, O Relly, Sebastopol, CA, 00. [1] L. B. Jang, S. C. Lew, Improvng Throughput and Farness by Reducng Exposed and Hdden Nodes n 80.11 Networks, IEEE Trans. on Moble Computng, vol. 7, no. 1, Jan. 008. [13] L. B. Jang, S. C. Lew, Hdden-node removal and ts applcaton n cellular WF networks, IEEE Trans. On Vehcular Technology, pp. 641-654, Sep. 007. [14] Clque (graph theory), http://en.wkpeda.org/wk/clque_%8 graph _theory%9. [15] L. Ca, X. Shen, J. W. Mark, L. Ca, Y. Xan, Voce capacty analyss of WLAN wth mbalanced traffc, Proceedngs of Qshne 05, Aug. 005 [16] Maxmal Clque, http://en.wkpeda.org/wk/ Maxmal_clque. [17] J. Snow, W. Feng and W. Feng, Implementng a low power TDMA protocol over 80.11, Proc. of IEEE WCNC 05, vol. 1, pp. 75-80, Mar. 005. [18] S. Sharma, K. Gopalan, N. Zhu, G. Peng, P. De, and T.-C. Chueh., Implementaton experences of bandwdth guarantee on a wreless LAN, Proc. ACM/SPIE Multmeda Computng and Networkng, Jan. 00 [19] T. Chueh and C. Venkatraman, Desgn, mplementaton, and evaluaton of a software-based real-tme Ethernet protocol, ACM SIGCOMM Computer Communcaton Revew, vol. 5, no. 4, pp. 7-37, 1995. [0] K. M. Svalngam, J. C. Chen, P. Agrawal and M. B. Srvastava, Desgn and analyss of low-power access protocols for wreless and moble ATM networks, Wreless Networks, vol. 6 no. 1, pp. 73-87, Feb. 000. [1] F. N. Al, et al., Dstrbuted and Adaptve TDMA Algorthms for Multple-Hop Moble Networks, IEEE MILCOM 0, pp. 546-551, Oct. 00. [] A. Kanzak, T. Hara, S. Nsho, An Adaptve TDMA Slot Assgnment Protocol n Ad Hoc Sensor Networks, ACM SAC 05, Santa Fe, USA, Mar. 005. [3] Z. Ca, M. Lu, SNDR: a new medum access control for multchannel ad hoc networks, IEEE VTC 00, Tokyo, Japan, May 00. [4] D. J. A. Welsh and M. B. Powell, An upper bound for the chromatc number of a graph and ts applcaton to tmetablng problems, The Computer Journal 10, pp.85-86, 1967. [5] IEEE Std. 80.11-1999, IEEE Standard for Wreless LAN Medum Access Control (MAC) and Physcal Layer (PHY) specfcatons, ISO/IEC 8 80-11: 1999(E), Aug. 1999. An Chan receved the B.Eng and M.Phl degrees n Informaton Engneerng from The Chnese Unversty of Hong Kong, Hong Kong n 005 and 007 respectvely. He s currently workng toward a Ph.D degree n the Department of Computer Scence at the Unversty of Calforna, Davs. Hs research nterests are n QoS over wreless network and advanced IEEE 80.11-lke mult-access protocols. He s a graduate student member of IEEE. Soung Chang Lew receved hs S.B., S.M., E.E., and Ph.D. degrees from the Massachusetts Insttute of Technology. From March 1988 to July 1993, Soung was at Bellcore (now Telcorda), New Jersey, where he engaged n Broadband Network Research. Soung s currently Professor and Charman of the Department of Informaton Engneerng, the Chnese Unversty of Hong Kong. Soung s current research nterests focus on wreless networkng. Recently, Soung and hs student won the best paper awards n the 1st IEEE Internatonal Conference on Moble Ad-hoc and Sensor Systems (IEEE MASS 004) the 4th IEEE Internatonal Workshop on Wreless Local Network (IEEE WLN 004). Separately, TCP Veno, a verson of TCP to mprove ts performance over wreless networks proposed by Soung and hs student, has been ncorporated nto a recent release of Lnux OS. Publcatons of Soung can be found n www.e.cuhk.edu.hk/soung.