Exploiting Idle Communication Power to Improve Wireless Network Performance and Energy Efficiency

Transcription

1 Explotng Idle Communcaton ower to Improve Wreless Network erformance and Energy Effcency Le Guo, Xaonng Dng, Hanng Wang, un L, Songqng Chen 3, Xaodong Zhang Oho State Unversty College of Wllam and Mary 3 George Mason Unversty Columbus, OH 43, USA Wllamsburg, VA 387, USA Farfax, VA 3, USA {lguo,dngxn,zhang}@cse.oho-state.edu {hnw,lqun}@cs.wm.edu sqchen@cs.gmu.edu Abstract As a famly of wreless local area network (WLAN) protocols between physcal layer and hgher-layer protocols, IEEE 8. has to accommodate the features and requrements of both ends. However, current practce has addressed the problems separately and s far from beng satsfactory. On one end, due to varyng channel condtons, WLANs have to provde multple data channel rates to support varous sgnal qualtes. A low channel rate staton not only suffers low throughput, but also sgnfcantly degrades the throughput of other statons. On the other end, TC s not energy effcent runnng on 8.. In a TC sesson, a wreless network nterface (WNI) has to stay awake to generate tmely acknowledgments, and hence, the energy s wasted by channel lstenng durng dle awake tme. In ths paper, consderng the needs of both ends, we utlze the dle communcaton power of the WNI to mprove the throughput and energy effcency of statons n WLANs wth multple channel rates. We characterze the energy effcency as energy per bt, nstead of energy per second. Based on modelng and analyss, we propose a data forwardng mechansm and an energy-aware channel allocaton mechansm. In such a system, a hgh channel rate staton relays data frames between ts neghborng statons wth low channel rates and the Access ont, mprovng ther throughput and energy effcency. Dfferent from tradtonal relayng approaches, our scheme compensates for the energy consumpton for data forwardng. The forwardng staton obtans addtonal channel access tme from ts benefcares, leadng to the ncrease of ts own throughput wthout compromsng ts energy effcency. We mplement a prototype of our proposed system and evaluate t through extensve experments. Our results show sgnfcant performance mprovements for both low and hgh channel rate statons. I. INTRODUCTION Moble devces are usually drven by battery power. Due to lmted battery capacty, t s essental to reduce power consumpton of moble devces wthout degradng ther performance. In moble communcatons, wreless network nterfaces (WNIs) consume a sgnfcant porton of energy. For nstance, the energy consumed by WNIs can account for more than 5% of the energy consumpton n handheld computers and up to % n laptop computers [4], [8]. As shown n [6], the energy consumpton of WNIs s domnated by the dle tme of WNIs, nstead of the amount of transferred data. To save energy n wreless devces, the basc prncple s to put the WNI nto sleep mode when t s dle, e.g., IEEE 8. power savng mechansm []. Nonetheless, due to the overhead of mode swtchng and lagged data recepton, frequent wakng up and sleepng of WNIs may result n serous performance degradaton and may even ncrease overall energy consumpton n moble devces [4], [6]. Furthermore, to mprove throughput and reduce response tme of clents, WNIs should always stay awake n TC sessons to generate tmely TC acknowledgments [6]. The attempt to sleep, whch nduces a delay n the generaton of ACKs, wll adversely affect TC throughput. In other words, durng an ongong TC sesson, the WNI has to be always actve. Thus, a sgnfcant porton of power s wasted on channel lstenng, whch we call the dle communcaton power of a staton. In addton to battery power, moble devces are very susceptble to physcal sgnal qualty degradatons such as fadng, attenuaton, and nterference. Due to varyng channel condtons, wreless local area networks (WLANs) have to provde multple data channel rates to support varous sgnal qualtes, such as IEEE 8.a (6-54 Mbps, 8 levels) and IEEE 8.b (- Mbps, 4 levels). The basc IEEE 8. channel access method guarantees an equal channel access probablty n the long-term to all statons. Snce a low channel rate staton takes a much longer tme to receve or transmt one data frame, t occupes a longer channel access tme and penalzes statons wth hgh channel rates. Therefore, low channel rate statons not only suffer low throughput themselves, but also sgnfcantly degrade throughput of other statons, and thus the entre WLAN [9]. To address ths performance anomaly n mult-rate WLANs, a tme-based farness channel access method [7] has been proposed, n whch each staton equally occupes the channel wth other statons, regardless of channel rates. However, whle the tme-based scheme protects hgh channel rate statons from dramatc performance degradaton, t aggravates the throughput of statons wth low channel rates. In ths paper, we utlze dle communcaton power to mprove throughput and energy effcency of moble statons n mult-rate WLANs. Instead of smply measurng the energy consumed on WNIs per second, we characterze the energy effcency n wreless communcatons as energy per bt, whch reflects the actual performance demands that users care about. Snce the WNI of a moble staton wth a hgh channel rate s dle n most tme, t can forward data frames as a proxy between ts neghborng statons wth low channel rates and the Access ont. Due to proxmty, the channel rates between

2 the proxy and ts clents are much hgher than those between the clents and the Access ont. Therefore, wth the proxy forwardng mechansm, we can sgnfcantly mprove the network performance and energy effcency of moble clents. Because the proxy consumes addtonal energy for data forwardng, we propose an energy-aware, n whch the proxy obtans addtonal channel occupancy tme from ts clents, resultng n the ncrease of ts own throughput wthout loss of ts energy effcency. Under such an ncentve mechansm, the forwardng servce s proftable and thus becomes a resource that statons want to compete for, whch s dfferent from prevous mult-hop routng algorthms n ad hoc networks. To ensure the farness of ths competton, we propose an aucton-based mechansm for proxy selecton. Based on the proposed mathematcal model, we analyze performance gans of proxes and clents n WLANs wth the support of mult-hop relay and channel tme allocaton. The analytcal results gve theoretcal bounds of performance gans under dfferent crcumstances. Accordng to the theoretcal gudelne, we elaborate the system desgn, whch conssts of three components: () a proxy selecton algorthm to choose relay statons for low channel rate statons; () a mult-hop forwardng algorthm to provde relable communcaton at the MAC layer and coordnate ntermedate statons along a forwardng path; (3) a token-based, energy-aware channel allocaton algorthm to provde channel occupancy tme compensaton to forwardng statons under tme-based farness and max-mn farness. To evaluate the proposed system, we mplement a prototype and conduct extensve experments on our testbed. Our expermental results show that ntegratng the proxy forwardng and energy-aware s, hgh channel rate moble statons (proxes) not only sgnfcantly mprove the network performance and energy effcency of low channel rate moble statons (clents), but also ncrease ther own throughput and the aggregate throughput of the entre WLAN, wthout compromsng ther energy effcency. Compared wth tmebased farness schedulng, the clent and proxy throughput can be mproved by 38% and by 3% respectvely, and the aggregate throughput of the entre WLAN can be mproved by 79%. The remander of ths paper s organzed as follows. Secton II surveys related work. Secton III dscusses our motvaton. Secton IV descrbes our system model and performance metrcs. Secton V presents the proxy forwardng and energyaware token rewardng mechansms. Secton VI detals our system desgn. We mplement a prototype of the proposed system and valdate ts effcacy on our testbed n Secton VII, and make concludng remarks n Secton VIII. II. RELATED WORK Most current WLANs support multple channel rates for moble statons wth dfferent sgnal qualtes. In outdoor WLANs, rado sgnal strength attenuates rapdly when the propagaton dstance ncreases. Studes [5], [7] have shown that rate dversty s prevalent n many ndoor WLANs and exsts even n a small room, because of the sgnal strength dversty caused by nose, nterference, mult-path, and user moblty. In [7], the authors also showed that wreless channels are often saturated due to channel contenton among dfferent users. Heusse et al. [9] dentfed a performance anomaly n 8.b that supports four dfferent channel rates. A moble staton transmttng at Mbps degrades the throughput of statons wth hgh channel rates (e.g., Mbps) down below Mbps. The man reason s that a moble staton wth lower channel rate takes much longer tme to transmt or receve a data frame, and hence, t occupes much more channel tme than hgher channel rate statons. To address ths anomaly, Tan and Guttag [7] proposed a tme-based farness schedulng algorthm n mult-rate WLANs. In ther algorthm, channel access tme s equally allocated among all statons wth dfferent channel rates. Thus, hgh channel rate statons are shelded from throughput degradaton, but the performance of low channel rate statons s decreased. IEEE 8. supports a power savng mechansm []. When a moble staton has no communcaton workload, t may swtch to power savng mode and notfy the Access ont to buffer ncomng data for t durng ts sleepng tme. In 8. WLANs, the Access ont perodcally broadcasts beacon messages so that moble statons can synchronze ther clocks. In each beacon message, the Access ont also transmts a traffc ndcaton map, whch contans a lst of statons that have data frames buffered at the Access ont. A moble staton n power savng mode perodcally wakes up and lstens to the beacon message. If there are data frames buffered at the Access ont for t, the staton polls the Access ont, and then the Access ont transmts the data frames to the staton. Afterwards the staton returns to sleep mode agan. IEEE 8. power savng mode may sgnfcantly degrade the network performance n TC [6] or RC [4]. Ths s because t ncreases the round-trp-tme (RTT) to the beacon nterval (about ms), whch s much greater than a typcal end-to-end RTT over the Internet. In [6], the authors demonstrated the performance degradaton of Web access caused by power savng mode, and proposed a bounded slowdown protocol to resolve the problem by adaptng the sleep and awake duratons based on the predcton of network actvtes. Anand et al. [4] have shown the performance degradaton of UD-based RC calls caused by power savng mode, and presented a self-tunng power management approach to adaptng the behavors of a moble staton to the access patterns and ntents of ts applcatons. Note that these solutons are orthogonal to our dle communcaton power explotaton, and can be ntegrated wth our proposed schemes for better network performance and power savngs. Explotng spatal reuse n cellular networks, Hseh and Svakumar [] have studed mult-hop ad hoc models to mprove network throughput and reduce energy consumpton for statons wth poor sgnal qualtes. However, spatal reuse s nfeasble n WLANs because of the channel overlappng problem. In [], a unfed cellular and ad-hoc network arch-

3 tecture has been presented, usng both a 3G cellular network nterface and a 8. network nterface. In [9], a relayenabled MAC protocol s proposed for ad hoc networks. In [9], the authors proposed a mult-hop WLAN archtecture and demonstrated ts benefts to wreless clents. However, none of these solutons can provde effectve ncentve mechansms to encourage statons to relay data for other statons. In contrast, our approach quanttatvely compensates proxy statons by rewardng them wth addtonal channel occupancy tme to mprove ther own throughput wthout compromsng energy effcency. To mprove TC performance n the wreless WAN envronment, nstead of WLAN, a proxy-based TC RISM [4] has been proposed, n whch mult-homed moble statons utlze the dversty of wreless WAN connectons for maskng adverse effects upon network performance. Besdes the closely related work mentoned above, extensve research has been conducted to nvestgate network performance and power consumpton n moble systems, such as [8], [5], and [6]. Moreover, a varety of transport-layer enhancements have also been developed to mprove networkng performance over wreless lnks, such as [5], [6], [7], [], [], [3], [7], [], [4], and [8]. III. MOTIVATION AND RATIONALE In ths secton, we descrbe the ratonale behnd the proposed data forwardng and channel access tme compensaton mechansms. We have two observatons. Frst, a moble staton nvolved n a TC sesson has to stay awake to generate tmely acknowledgments (ACKs) upon data arrvals. Channel lstenng consumes a consderable amount of energy. Second, a moble staton wth a low channel rate sgnfcantly reduces the throughput of statons wth hgh channel rates and plummets the aggregate throughput of the entre WLAN. One soluton to these problems would be to recrut moble statons wth hgh channel rates to harvest ther dle tme and forward data frames for the moble statons wth low channel rates. A low channel rate staton mproves ts throughput va a much hgher channel-rate path. However, a hgh channel rate staton has to consume extra energy on forwardng data frames for the low channel rate statons, whch t may be unwllng to do. Instead of forcng the hgh channel rate statons to sacrfce ther energy for data relay, the low channel rate statons should compensate a certan amount of tme slots to the hgh channel rate statons, and hence, at least the hgh channel rate statons wll not be penalzed by beng helpful. Intally, each moble staton should be assgned the same amount of channel access tme for data communcaton, followng the rule of tme-based farness [7]. In our scheme, the allotted tme for each staton can be traded for hgher throughput. A moble staton can mprove ts throughput ether by obtanng more tme slots for ts own communcaton or by ncreasng the channel rate at whch ts data are transmtted. If a hgh channel rate staton obtans extra tme slots from low channel rate statons, and a low channel rate staton ncreases ts data transmsson rate through a hgh channel-rate path composed of proxes; then t wll be a wn-wn scenaro. A moble staton always desres a hgh throughput and low energy consumpton. The effcency of energy utlzaton needs a lttle more careful consderaton. Energy consumpton can be expressed as energy consumed per unt tme, or energy consumed per data volume. A fxed data transmsson rate for a moble staton gves us an lluson that a user cares about energy consumpton per unt tme, whch s not always true. We beleve that the user actually cares more about how much energy consumed for a certan amount of data communcated, because the WNI can be put nto sleep mode or turned off when t has no communcaton workload. We defne the energy utlty of the WNI of a moble staton as the average number of effectve bts transmtted/receved per unt energy when the power s on. Thus, the best way to save energy s to reduce the energy cost of every effectve bt or ncrease the energy utlty, and turn off the WNI when the communcaton sesson termnates. To encourage a hgh channel rate staton to relay data for a low channel rate staton, ts energy utlty should not be reduced. A WNI can work n three modes wth dfferent power consumpton levels: transmsson, recevng/lstenng, and sleep mode. The power consumpton of transmsson mode s usually much hgher than that of recevng/lstenng mode. Thus, the energy utlty of a hgh channel rate staton wll be lowered f t relays for a low channel rate staton wthout any compensaton. However, f the low channel rate staton contrbutes a fracton of ts allotted tme slots to the proxy staton, the proxy staton can use these bonus tme slots for ts own communcaton, leadng to the ncrease of ts throughput and the decrease of ts WNI workng tme. As a result, although the proxy staton spends extra energy for the data forwardng servce, ts energy utlty can reman ntact or even ncrease. IV. SYSTEM MODEL AND ERFORMANCE METRICS We now focus on network descrpton and basc notatons before we proceed to dscuss protocol desgn. The WLAN n consderaton s composed of an Access ont (A), S, and n (n ) moble statons (denoted as statons n the remander of ths paper), S, S,..., S n. The rado channel s shared by all statons and the Access ont. Two statons S and S j can communcate wth each other at a channel rate R,j ( j and, j n). Specfcally, each staton S ( n) can communcate wth the A wth channel rate R,, and we denote R, as R for smplcty. Let t be the power consumpton (energy per second) of a staton n the transmsson mode, and r be the power consumpton of a staton n the lstenng or data recevng mode. Assume t = α r (α > ). We further assume the fracton of allocated channel occupancy tme of staton S s t, n whch the fracton for data transmttng s f ( f ). For example, the typcal power consumpton of the Csco Aronet 35 seres WNI s 45 ma at transmsson mode, 7 ma at recevng/lstenng mode, and 5 ma at sleep mode (all under 5V DC) [].

4 symbol (S ) T(S ) E(S ) R,j f t x j,k yj U(S ) gt (S ) ge (S ) g T (S ) g E (S ) TABLE I SYMBOLS AND NOTATIONS meanng and unt power consumpton of staton S (Joule/sec) throughput of staton S (bt/sec) energy utlty of staton S (bt/joule) the channel rate between staton S and S j (bt/sec) the fracton of outgong traffc n S s workload the allocated tme of S the fracton of channel tme durng whch the traffc of S s forwarded between S j and S k the fracton of channel tme that S rewards S j utlzaton of allocated tme of staton S the throughput gan when S has no clents the energy utlty gan when S has no clents the throughput gan of S the energy utlty gan of S In tme-based farness schedulng [7], each staton s assgned the same fracton of channel tme. Thus, t = t = n ( n), and we also have the bound < t. Based on the energy consumpton of a staton S per unt tme, (S ), we defne two performance metrcs for a wreless staton as follows: Throughput, T(S ), the number of effectve bts a staton transmts and receves per unt tme; Energy utlty, E(S ), the average number of effectve bts per unt energy. That s, E(S ) = T(S) (S ). Accordng to the assumptons of our model, we have (S ) = t t f + r ( t f ) = r ( + (α )t f ), T(S ) = R t, E(S ) = (α )f + t R r, (IV.) where n. We can compare the orgnal tme-based farness scheme wth our new scheme by consderng the performance gan n terms of throughput and energy utlty: { gt (S ) = T (S ) T(S, ) g E (S ) = E (S ) E(S, (IV.) ) where (S ), T(S ), E(S ) and (S ), T (S ), E (S ), are the power consumpton, throughput, and energy utlty of a staton S before and after a forwardng servce t provdes/receves, respectvely. Table I lsts the notatons that are used n the followng analyss. V. CHANNEL ALLOCATION FOR MULTI-HO FORWARDING In ths secton we nvestgate the channel allocaton for supportng mult-hop forwardng. More specfcally, how much tme a low rate staton has to offer the hgh rate staton for the forwardng servce so that the latter wll not be penalzed. We analyze a smple one-hop case frst, and then extend the one-hop relay to the general case of mult-hop relay. The bts for MAC level retransmsson and the forwardng data for other statons are not counted as effectve bts. Low rate chanel Hgh rate chanel S A Fg.. S p proxy S p forwards data for S q A. Channel Occupancy Tme Allocaton S q clent Assumng that the tme-based farness schedulng s enabled, each staton s assgned an equal fracton of channel tme n unts of tme slot. In such a WLAN, a staton S p that can communcate wth A at a hgh channel rate can work as the proxy staton for a staton S q that can only communcate wth A at low channel rates, as long as the proxy and clent can communcate wth each other at hgh channel rates, as shown n Fgure. To enable such a servce, the tme slots used for data forwardng should come from the tme slots of the clent statons. Meanwhle, snce transmttng data for clents consumes ts energy, the proxy staton should be rewarded addtonal tme slots from ts clent statons for compensaton. We defne the fracton of channel tme that a clent S q rewards ts proxy S p to keep the energy utlty of S p unchanged as the cost prce (or valuaton) of the forwardng servce, denoted as cost(p, q). We defne the fracton of channel tme that a staton s assgned by the tme-based farness schedulng as the assgned tme of the staton, and the fracton of channel tme that a staton can use for ts own communcaton as the effectve tme of the staton. We also defne the fracton of channel tme that a clent rewards each of ts proxes as ts rewardng tme to the proxy or the rewarded tme of that proxy. The effectve tme of a proxy s ts assgned tme plus all rewarded tme from ts clents. The effectve tme of a clent s ts assgned tme subtractng the fracton of channel tme t rewards ts proxes and the fracton of channel tme for ts data relayng (transmttng or recevng) between the A and the mmedate proxy of the clent (relayng tme). We further defne the sum of a staton s assgned tme accordng to tme-based farness schedulng and ts rewarded tme from ts clents as the allocated tme of the staton, whch can be used for ts own communcaton or be rewarded to ts proxes. Therefore, we defne the utlzaton of the allocated tme of a staton S, U(S ), as the rato of ts effectve tme to ts allocated tme. B. erformance Gan Analyss for One-hop Relay Frst, we consder one clent and one proxy for smplcty. Assume clent staton S q s relayed by proxy staton S p. The assgned tme of S q should be dvded nto three peces: t q = t = x,p + x p,q + y q p, (V.3) where x,p s the fracton of channel tme used for data relayng between A (S ) and proxy staton S p (relayng tme), x p,q s the fracton of channel tme that clent staton S q s transmttng/recevng data to the proxy staton (effectve

5 S p S q assgned tme of S p t x, p x p, q y p q assgned tme of S q Fg.. effectve tme of S p Channel tme allocaton t y p q x, p relayng tme for S q x p, q effectve tme of S q tme), and yp q s the fracton of channel tme that the clent staton compensates S p (rewardng tme). The utlzaton of S q s allocated tme s U(S q ) = xp,q t. The effectve tme of S p s t p = t p + y q p = t + yq p, (V.4) where t p s ts assgned tme and yp q s ts rewarded tme from clent S q. The utlzaton of S p s allocated tme s snce t can use all ts assgned tme and rewarded tme for ts own communcaton. Fgure shows the channel tme allocaton n one-hop proxy forwardng. Lemma : In one-hop forwardng, the allocated tme utlzaton, rewardng tme, throughput gan and energy utlty gan of a clent S q when t pays the cost prce to ts proxy S p for the forwardng servce are R U(S q ) =,p R,,p+R p,q+(α ) t[f qr p,q+( f q)r,p] yp q = ( t) U(S q )R p,q (α )( fq R,p + fq R p,q ), g T (S q ) = Rp,q R,q U(S q ), g E (S q ) = Rp,q (α ) tf R,q U(S q ) q+ U(S. q)(α ) tf q+ roof: Two constrants dctate how much tme a low rate staton has to offer to a hgh rate staton: () every clent staton allocates suffcent tme for the transmsson and forwardng of ts data; () the energy utlty of the hgh rate staton remans the same. Frst, we have T (S q ) = x,p R,p = x p,q R p,q, (V.5) whch mples that the flow rate n each hop along the forwardng path of clent S q are equal. Second, the energy utlty of the proxy s unchanged, that s, the cost prce of S p servng S q s the rewardng tme of S q to keep the energy utlty of S p unchanged E(S p ) = E (S p ). (V.6) Equaton IV. gves the power consumpton, throughput and energy utlty of S p when t has no clents. Denote the power consumpton, throughput and energy utlty of S p when S p serves clent S q as (S p ), T (S p ), and E (S p ), respectvely, we have (S p ) = r ( + (α )t f p), T (S p ) = R(S p )( t + yp q), E (S p ) = (S p) T (S, p) (V.7) where t f p = f p ( t+yp)+f q q x,p +( f q )x p,q s the total tme of proxy S p used for data transmsson. In t f p, f p ( t + yp) q s the tme that S p transmts ts own upstream workload to A, f q x,p s the tme that S p forwards the upstream workload of S q to A, and x p,q s the tme that S p forwards the downstream workload of S q to S q. Resolvng Equatons V.3, V.5, and V.6, we have T (S q ) = R,p + Rp,q U(S q ) = xp,q t = T (S q) R p,q t = t fq +(α ) t( R + fq,p Rp,q ), R p,q R,p+R p,q+(α ) t[( f q)r p,q+f qr,p], yp q = t p (α )(f q x,p + ( f q )x p,q ) = t(α )T (S q )( fq R,p + fq R p,q ), (V.8) where yp q = cost(p, q). Accordng to Equaton IV., for clent staton S q, we have { gt (S q ) = Rp,q R,q U(S q ), g E (S q ) = Rp,q R,q U(S q ) (α ) tf q+ U(S. q)(α ) tf q+ (V.9) U(S q ) and g T (S q ) ncrease wth the ncrease n the number of statons (the decrease of t) n the WLAN. We have R U(S q ) < p,q R,p+R p,q and g T (S q ) < Rp,q R p,q R,q R,p+R p,q. Relayng s only useful when the throughput gan g T (S q ) >. Snce U(S q ) <, f q, by examnng Equaton V.9, we have g E (S q ) g T (S q ). That s, relayng can always ncrease the energy utlty of a clent staton as long as ts throughput can be mproved. For a specal case when R,p = R p,q, we have yp q = (α ) t +(α ) t, < yq p (α+3), T tr (S q ) =,p +(α ) t, < T (S q ) R,p α+3, U(S q ) = +(α ) t, α+3 U(S q) <, (V.) g T (S q ) = g E (S q ) = R,p R,q, +(α ) t R,p α+3 +(α ) tf q +(α ) t(+f q) R,q g T (S q ) < R,p R,q. R,p R,q, A proxy staton can serve multple clents at the same tme, and these clent statons may have dfferent channel rates and dfferent data transmttng/recevng ratos. We have the followng lemma. Lemma : Assume staton S p provdes forwardng servces to k clent statons, S q, S q,..., S qk (k > ), and these clent statons ndependently contrbute ther rewardng tme to S p to keep the energy utlty of S p unchanged, we have U(S p ) =, g T (S p ) = + (α ) k = T (S q )( fq R,p + fq R p,q ), g E (S p ) =, where T (S q ) s the throughput of clent S q ( k) when the forwardng servce s on. roof: It s easy to see that U(S p ) = and g E (S p ) =. Snce each clent rewards S p ndependently, smlar to the last

6 formula n Equaton V.8, we have t p = (α )(fq x,p+( fq )xp,q ) y q p =... = (α )(fq k x,p+( fq k )xp,q k ) y q k, p k = [(fq x,p+( fq )xp,q k )] t p = +(α ) t k p+ = yq p The effectve tme of S p s t p = t p + k have g T (S p ) = T (S p) T(S = t p p) t p = + k = yq p t p = yq. p. Thus, we = + (α ) t k = T (S q )( fq R,p + fq R p,q ). In case R,p = R p,q ( k), we have k t g T (S p ) = + (α ) + (α ) t. (V.) Snce k t = k n <, g T(S p ) s bounded by < g T (S p ) < α +. (V.) C. A Generc Analyss for Channel Allocaton n Mult-hop Forwardng A staton S that s relayed by other statons can stll work as the proxy for statons wth even lower channel rates, and gets rewarded tme from ts clents. However, only a fracton of ts rewarded tme can be used for ts own communcaton, snce S also needs to reward ts relayng statons. We consder the relay chan S S S S startng from the A (S ). In order for S to relay data for S, S has to keep ts energy utlty unchanged. After S decdes to relay data for S, S wll have a hgher energy utlty than before. S would lke to keep ths new energy utlty unchanged when t decdes to relay for, and so on. The followng Lemma descrbes the performance gan of a staton n such scenaros. The proof bascally formalzes the above process. Denote the throughput gan and energy utlty gan when S has no clents as gt (S ) and ge (S ), respectvely. We have the followng lemma. Lemma 3: Assume each staton has at most one mmedate relayng staton n a WLAN, and each staton rewards ts relayng statons ndependently to keep ther energy utltes unchanged. For staton S that s relayed by ( ) statons along the path S S... S S, and S has m ndrect or drect clents (S q, S q,..., S qm ), we have { g T (S ) = R, R, U(S ), ge (S ) = R, (α ) tf R, U(S ) + U(S, )(α ) tf + where U(S ) = and {, +R, j= [ f R +(α ) t( j,j R + f j,j R )] j,j+ m g T (S ) = gt (S j= )( + yq j t ), g E (S ) = ge (S ), where y qj = t(α )T (S qj )( fq j R, + fq j R,j ), T (S qj ) s the throughput of S qj when t s forwarded by S, and S j s the next hop staton of S to reach S qj. roof: For staton S ( > ) that s relayed by statons S,..., S, we have t = (x, +...+x, )+x, +(y +...+y ). (V.3) The flow rate of S s own traffc n each hop along the forwardng path s equal, so we have T (S ) = x,r, =... = x, R, = x,r,. (V.4) For a relayng staton of S, S j ( < j < ), when S j has no clents, we have (S j ) = t tf j U(S j ) + r ( tf j ) = r [ + (α ) tf j U(S j )], T(S j ) = R(S j ) tu(s j ), (V.5) where U(S j ) = when S j has no proxy (j = ), and U(S j ) < when S j s relayed by other statons ( < j < ). When S j serves staton S j+,..., S, we have { (S j ) = r [ + (α )t f j ], T (S j ) = R(S j )( t + l=j+ yj l )U(S (V.6) j), where t f j = f j( t + l=j+ yj l )U(S j) + l=j+ f lx l j,j + l=j+ ( f l)x l j,j+. In tf j, f j( t+ l=j+ yj l )U(S j) s the tme used by S j to transmt ts own workload to S j, f l x l j,j s the tme used by S j to transmt the upstream workload of S l to S j, and ( f l )x l j,j+ s the tme used by S j to transmt the downstream workload of S l to S j+. Consderng the energy utlty of S j, we have E(S j ) = R(Sj) E (S j ) = tu(s j) r +(α ) tf, ju(s j) R(Sj) ( t+ l=j+ yj l )U(Sj) r. +(α )t f j The energy utlty of S j should be unchanged, that s, E(S j ) = E (S j ). By substtutng E(S j ) and E (S j ), we have (α )f j + tu(s j) = (α )f j + +(α ) l=j+ (f lx l j,j +( f l)x l j,j+ ) ( t+, l=j+ yj l )U(Sj) Smplfyng the equaton, we have t = (α ) l=j+ (f lx l j,j +( f l)x l j,j+ ). l=j+ yj l Snce each staton S l (j + l ) rewards tme slots to S j ndependently, we get Thus, we have t = (α )(f lx l j,j + ( f l)x l j,j+ ). y j l = t(α )(f l x l j,j + ( f l)x l j,j+ ) = t(α )T f (S l )( l R j,j + f l R j,j+ ), y j l (V.7)

7 where T (S l ) s the throughput of S l when t s served by S j and T (S l ) = R l,l t l U(S l ), where U(S l ) s the allocated tme utlzaton of S l. When S has no clents, we have t = t. Consderng Equaton V.3, V.4, and V.7, for staton S, we have Access ont moble staton S S staton S S S 4 S 5 S 6 S 7 relayed by channel rate R, R, R, 3 R, 4 R, 5 R, 6 R, 7 U(S ) = T (S ) R,t =. +R, j= [ f R +(α ) t( j,j R + f j,j R )] j,j+ (V.8) Accordngly, we get gt (S ) = T (S ) T(S = R,tU(S) ) R, t = R, R, U(S ), ge (S ) = E (S ) E(S = R, ) R, U(S ) (S) (S ) = R, (α ) tf R, U(S ) + U(S. )(α ) tf + (V.9) When S has m clents S q,..., S qm, snce each clent rewards S tme slots ndependently, the throughput becomes T (S ) = U(S )R, ( t + m j= yqj ). Thus the performance gan s { g T (S ) = T m (S ) T(S ) = gt (S j= )( + yq j t ), g E (S ) = ge (S ), where y qj follows Equaton V.7. The above analyss of the channel allocaton for one-hop and mult-hop forwardng shows the performance gans of low channel rate clents and hgh channel rate proxes. Specfcally, we show that our proposed scheme can even ncrease the proxy s throughput wthout compromsng ts energy utlty, provdng a strong ncentve for beng a proxy. VI. SYSTEM DESIGN In ths secton, we descrbe the system desgn of the multhop forwardng servce. The proposed system conssts of three major components: a proxy selecton algorthm, a token-based energy-aware channel schedulng algorthm, and a mult-hop forwardng algorthm. The proxy selecton algorthm runs on A, choosng relay proxes for statons wth low channel rates. The energy-aware channel schedulng algorthm also runs on A, arbtratng channel tme allocaton and ensurng tmebased and max-mn farness among statons. The mult-hop forwardng algorthm s a dstrbuted algorthm runnng on both A and each staton, n order to coordnate ntermedate statons along the forwardng path and provde relable communcaton at the MAC layer. The three algorthms work together to enable the data forwardng among statons n a WLAN. As shown n Fgure 3, statons n the WLAN are organzed nto a tree rooted at the A for the mult-hop forwardng servce. Each non-root node of the tree represents a staton, and the weght of each edge represents the channel rate between two nodes. The A (root) mantans the topology and edge weghts of the forwardng tree. Each staton mantans the nformaton about ts chldren and predecessors, and the weght of each edge along the path. Note that the heght of the 3 S Forwardng Table S 5 self R, 3 R 3, 5 R 5, 6 S 4 hop staton channel rate Fg. 3. S 6 S 5 S 7 hop staton channel rate 3 self S 6 S 7 Mult-hop forwardng structure forwardng tree should be small (typcally two or three n 8.b). The man reason for ths s that the data forwardng along each hop requres the occupancy of channel resources (spatal reuse s dffcult n WLANs). Wth the ncrease n the number of forwardng hops, the mprovement of a clent s throughput decreases rapdly. Moreover, due to the possble moblty of statons, t s much easer to mantan a short tree than a tall one. A. roxy Selecton and Assocaton Wth the channel tme compensaton, the forwardng servce s proftable and thus becomes a resource that statons want to compete for. Ths s dfferent from prevous mult-hop routng algorthms n ad hoc networks. To ensure the farness of ths competton, we propose an aucton-based mechansm for proxy selecton. Our proxy selecton algorthm runs on the A, whch works as the auctoneer. When a staton S q communcates wth the A at a low channel rate, t broadcasts a sequence of SF (search for proxy) messages wth dfferent channel rates. Upon recevng an SF, each hgh channel rate staton computes the expected throughput gan t can provde to S q and the cost prce based on Lemma 3, then bds for the forwardng servce wth the cost prce. Upon recevng an SF, the A collects the bds from all bdders wthn the bddng tme, and then selects the staton that can provde the largest throughput gan for S q as the proxy. A clent would always lke to pay less and get more, whle a proxy would always lke to beng pad more and serve less. In our mechansm desgn, the domnant strategy for all bdders the best strategy they can expect should be to bd wth the cost prce of ther servces. We use the second prce sealed bd aucton rule [3] to provde such a domnant strategy and fnsh the aucton n one bddng round. In ths mechansm, staton S q wll pay the proxy at the prce of the bdder who offers the second largest throughput gan (see our techncal report [8] for the detaled descrpton of the aucton mechansm). When the proxy s selected, the A sends (or pggybacks) the MAC address of the proxy and the correspondng prce to S q. Then S q sends a RFR (request for relay) message to the R, 3 R 3, 5 R 5, 6 R 5, 7

8 arrvng packets /r sec X arrvng tokens from A bucket depth = 3 /r sec 3/r sec to wreless channel Fg. 4. Token bucket: the A dstrbutes tokens n a rate r (one round per /r seconds) proxy, and the proxy acknowledges the request and reports to the A to commt the proxy assocaton. When the clent does not need data forwardng any longer, t sends a notfcaton to the A drectly through the low-rate channel to cancel the forwardng servce. Many hgh channel rate statons may compete wth each other to obtan more rewarded tme slots for mprovng ther own throughput. The A needs to balance the profts among proxy canddates that can provde the same forwardng servces n a WLAN. For example, f two statons can provde the same throughput gan for S q, the A should favor the staton wth less throughput than the proxy of S q. Other factors, such as the hstory of actvty and the moblty of the proxy canddates, may also be taken nto consderaton for proxy selecton. B. Channel Allocaton and Schedulng The channel schedulng and the forwardng coordnaton can be easly mplemented n 8. WLANs under CF (pont coordnaton functon) wth pollng MAC mechansm. However, most 8. commercal products only support DCF (dstrbuted coordnaton functon) MAC control. In what follows, we descrbe our system desgn for 8. WLAN under the DCF MAC mechansm. In the proposed system, the channel s allocated n unts of tme slot, same as the unt of staton s back-off tme for HY medum access (5 µs for FHSS and µs for DSSS). As shown n Fgure 4, the tme slot allocaton s performed by the A based on the token bucket model. Each staton s assgned a certan number of tokens for channel contenton. A staton competes for channel only when t has avalable tokens. At regular ntervals, the A evenly dstrbutes tokens to each staton, ensurng tme-based farness. When the bucket of a staton s full, the overflowng tokens are returned to A, and are re-dstrbuted equally to other statons for max-mn farness. The token bucket shapes the frame transmsson of a staton at a constant rate n the long run, whle allowng bursty frame transmsson of a staton n the short term. The tokens can be dstrbuted ndvdually or be pggybacked wthn the data/control frames to statons. A staton transmts data frames only when t has enough tokens. Smlarly, the A buffers data frames for statons wthout tokens, and postpones ther data transmsson to the next round of tme-slot allocaton. Thus, the number of tokens a staton holds determnes whether t s qualfed for channel competton. Meanwhle, channel contenton s far for those statons wth tokens. Therefore, the channel occupancy tme of a staton s dependent on the token allocaton scheme n the long term, although t s non-determnstc n the short term. We use the smlar method as that n [7] to compute the channel occupancy tme of a staton. For each staton, there are two token counters, one mantaned at the staton tself and the other at the A. Upon recevng/sendng a data frame from/to the A, the staton deducts the correspondng tokens from ts token counter. At the same tme, the A deducts the same number of tokens of that staton as well. In 8. protocol, the number of retres of a successfullytransmtted frame s ncluded n the frame header, so that the recever clearly knows t. However, current hardware does not return the number of retres when the frame s successfully transmtted. Thus, the sender cannot accurately compute the number of tokens used for data transmsson, and the two counters may be nconsstent. To mnmze ths effect, the recever pggybacks the number of tokens that are used for the last data transmsson of ts peer n the data frame, and the peer adjusts ts token counter accordngly. To smplfy token management, a proxy staton does not mantan token counters for ts clents. Once a clent assocates to the proxy, the tokens, ncludng those that the clent should reward ts proxy and those that are used to receve/forward data frames for the clent, are delvered to the proxy drectly by the A durng the token dstrbuton. Correspondngly, the same number of rewardng tokens s deducted from the token counter of the clent by the A. Once the clent cancels the forwardng servce, the proxy automatcally suspends the data forwardng at the next round of token dstrbuton, because the A wll no longer convey the clent s rewardng tokens. C. Mult-Hop Forwardng ) Basc Mechansm: To support mult-hop forwardng, each data frame s appended wth two felds ndcatng the source and destnaton MAC addresses of the frame, respectvely. Each staton mantans a forwardng table as shown n Fgure 3. Upon recevng a data frame, the staton compares the destnaton MAC address wth ts own MAC address. If they are dfferent, the staton looks up the MAC address for the next-hop staton n the forwardng table. Then t modfes the destnaton address of the frame header and forwards t to the next-hop staton. The forwardng table also records the uplnk channel rates of the staton s predecessors, n order to compute the cost prce of the forwardng servce, and the throughput gans ts clents can acheve. ) Forwardng ath Mantenance: The channel rates along the forwardng path and the one between the clent and the A may change wth the moblty of statons or sgnal nstablty. Furthermore, the forwardng path may even be broken, due to hardware falure, sgnal error (or nterference), and the moblty of proxy statons. To cope wth the possble change of channel rates, each clent perodcally re-evaluates

9 the forwardng servce. If the servce qualty s sgnfcantly degraded, t re-broadcasts SFs for a new proxy. 3) ower Management n Mult-hop Forwardng: Most exstng power savng solutons [4], [6] utlze heurstc algorthms to adapt the sleepng of a WNI wth ts network actvtes. When a staton has no network traffc, t wll stll be up for a whle before t goes to sleep based on the predcton of ts network actvty. The staton may also change ts wakng perod adaptvely to save energy consumpton on beacon lstenng. In our scheme, each staton has the flexblty to set ts own power savng polcy. Any staton that wants to sleep needs to send a request to the A, so that the A can buffer the ncomng data frames for t. The sleep request of a clent s drectly sent to the A at a low channel rate. When a proxy decdes to swtch to power savng mode, t notfes ts mmedate clents frst. If any chld of the proxy has clents, the notfcaton wll be propagated recursvely. Upon recevng the ACK from all ts clents, the proxy sends a request to the A, and shfts to power savng mode. Then, ts clents (mmedate or non-mmedate) search for new proxes. VII. IMLEMENTATION AND EXERIMENTAL EVALUATION Ths secton presents the prototype mplementaton of our proposed scheme and ts expermental evaluaton. Our purpose s twofold: () to demonstrate that our data forwardng mechansm s feasble under the framework of the current IEEE 8. protocol; and () to valdate ts effcacy n sgnfcantly mprovng the throughput and energy utlty for statons n the WLAN. A. rototype Implementaton We have mplemented a prototype of the proposed scheme and evaluated t on our testbed, whch ncludes an Access ont and sx moble statons. The A s a desktop C equpped wth a NetGear MA3 8.b CI wreless adaptor runnng Lnux kernel.4.. The moble statons are sx H laptop computers, each equpped wth a NetGear MA4 8.b CMCIA wreless adaptor runnng Lnux kernel.4.. One of the sx works as the proxy, the others work as the clents. All wreless adaptors n the A and moble statons use the Intersl rsm chpset. We have modfed the HostA Lnux drver for rsm/.5/3 [] as the drver of our Access ont. The A mantans the forwardng structure for each staton assocated wth t, as descrbed n Secton VI. The bddng tme for proxy selecton s set to 5 ms and the token dstrbuton nterval s set to ms. Each token denotes µs channel occupancy tme. To mplement the token dstrbuton, the HostA drver mantans the number of avalable tokens owned by each moble staton that s currently assocated wth the A. In each round of the token dstrbuton, the HostA drver frst evenly allocates tokens based on the number of statons, then transfers the rewardng tokens from each clent to ts proxy based on ther servce agreement. Fg. 5. rates throughput (Mbps) Mbps Mbps 5.5 Mbps Mbps channel rate The effectve bandwdth of 8.b WLAN under dfferent channel We have also modfed the ORNOCO Lnux drver.5rc for wreless cards [3] as the drver of our proxy and clent statons. Insde the drver, we have mplemented a smple mult-hop forwardng protocol. B. Expermental Evaluaton To evaluate the mplemented prototype, we conduct extensve experments on our testbed wth respect to FT-lke and Web-lke workload, respectvely. Due to page lmtatons, we only present the results of FT-lke workload (see our techncal report [8] for the detaled evaluaton of Web-lke workload). ) erformance Baselne Measurement: The deal channel rate of IEEE 8. WLAN cannot be acheved n realty, due to the overhead of control frames, nter-frame spaces, physcal and MAC layer headers, channel contenton, and possble data losses. Therefore, we frst measure the effectve throughput of a WNI as the baselne for performance comparson. In ths evaluaton, we frst set up a small 8.b WLAN that conssts of an A and a moble staton. We transfer a large fle (about GB) from the A to the staton, and measure the user level throughput under dfferent channel rates. Fgure 5 shows the effectve bandwdth of the 8.b WLAN under channel rates of Mbps, Mbps, 5.5 Mbps, and Mbps, respectvely. The hgher the channel rate, the less effcent the channel utlzaton. The reason s that all physcal layer headers are transmtted at the lowest channel rate n the 8.b protocol, n order to ensure that every staton can lsten to the channel for collson avodance. However, the dversty of user level throughput under dfferent channel rates s stll large enough to beneft low channel rate users through data forwardng. In WLANs wth more levels of channel rates such as 8.a, mult-hop data forwardng would have greater potental to mprove the system performance. ) Experments on FT-lke Workload: We mplement four channel allocaton protocols as lsted n Table II and compare ther throughput and energy utlty wth FT-lke data transmsson workload. In these schemes, DCF denotes the normal DCF MAC n a 8.b WLAN, and TBF denotes the tme-based farness channel contenton mechansm proposed n [7]. SFW denotes our proposed mechansm, meanng selfsh mult-hop forwardng, n whch the clent pays the

10 throughput (Mbps) energy utlty (Mb/J) 3 3 ( M) ( M) overall.35 staton to staton : M overall.93 overall 3.99 overall 3.8 ( M) ( M) overall.67 overall.46 overall.86 overall.76 (a) Channel rates between statons: -A, M; -A, M; -, M throughput (Mbps) energy utlty (Mb/J) ( M) (5.5 M) overall.5 staton to staton : M overall.3 overall.7 overall.63 ( M) (5.5 M) overall.63 overall. overall.8 overall.4 (b) Channel rates between statons: -A, M; -A, 5.5 M; -, M Fg. 6. The throughput and energy utlty of statons under dfferent s ( proxy and clent) Scheme DCF TBF SFW TBF-FW TABLE II CHANNEL ALLOCATION SCHEME Scheme Descrpton 8. DCF MAC (wthout data forwardng) tme-based farness schedulng (wthout data forwardng) selfsh forwardng under TBF schedulng data forwardng under TBF schedulng cost prce for the forwardng servce (because there s only one proxy n our testbed). In order to show the advantage of our proposed channel tme compensaton mechansm, we also mplement data forwardng under tme-based farness for comparson, called TBF-FW. In ths mechansm, each staton s assgned equal channel tme to ensure tme-based farness, and the proxy voluntarly forwards data for ts clents usng the channel tme of ts clents, wthout any tme slot rewarded. Note that ths s a phantom mechansm just for comparson, nether proposed nor mplemented before. In the experments, we smultaneously download a large fle from the HostA machne to the proxy and clent statons, respectvely. The throughput s measured by recordng the data volume transfered between each clent and ts proxy (or between the proxy and the A) under dfferent channel allocaton schemes. The energy consumpton on data transmsson s computed as the product of the data transmsson tme of physcal frames and the power consumpton of the wreless card n the transmttng mode (provded by the manufacturer). The energy consumpton on recevng/lstenng s computed n a smlar way. We conduct experments for the one-hop forwardng case, where the WLAN conssts of A, proxy (denoted by ), and multple clents (denoted by ) varyng from to 5. Assumng all clents have the same channel rate, there are eght possble combnatons for the data forwardng servce: the channel rate s M or M between -A, M between -A, and M between -; the channel rate s M or M between -A, 5.5 M between -A, and M between -; the channel rate s M or M between -A, M between -A, and 5.5 M between -; the channel rate s M or M between -A, 5.5 M between -A, and 5.5 M between -. Each experment s repeated three tmes. Fgures 6, 7, and 8 show the performance n a WLAN wth A, proxy, and, 3, and 5 clents, respectvely. Due to page lmtatons, only part of results are presented (other results are smlar). In the fgures, the number on the top of the bar group denotes the overall throughput (n Mbps) or the overall energy utlty (n Mb per Joule) of the proxy and clent statons n the WLAN. We also present the correspondng performance of DCF and TBF for comparsons. The performance of phantom TBF-FW s presented as whte bars. The results can be summarzed as follows. SFW has the hghest overall performance wth respect to both throughput and energy utlty, whle DCF has the worst overall performance. By enforcng tme-based farness, TBF mproves the performance of hgh channel rate statons but decreases the performance of low channel rate statons. TBF-FW mproves the throughput of low channel statons (clents) by data forwardng, but sgnfcantly decreases the energy utlty of the forwardng staton (proxy), whch the proxy s unwllng to do. Thus ths phantom scheme s not lkely to be feasble n practce. In contrast, n our proposed forwardng scheme, the proxy receves addtonal channel tme compensaton from ts clents, resultng n the mprovement of ts own throughput wthout decreasng ts energy utlty. The clent statons sacrfce a few channel tme tokens for the forwardng servce, but the overhead s mnor. For example, as shown n Fgure 7(a), the clent throughput of SFW s 38% hgher than that of DCF, more than tmes over that of TBF, and about 93% of that of TBF-FW, whle the proxy throughput of SFW s more than 5 tmes over that of DCF, and 3% hgher than those of TBF and TBF-FW. The proxy energy utlty of SFW s more than 4 tmes over that of DCF, and s same as that of TBF. On the other hand, compared wth SFW, the proxy energy utlty of TBF-FW s % lower than that of

11 throughput (Mbps) energy utlty (Mb/J) ( M) ( M) overall.98 staton to staton : M overall.85 overall 3.3 overall 3.7 ( M) ( M) overall.5 overall.46 overall.78 overall.75 (a) Channel rates between statons: -A, M; -A, M; -, M throughput (Mbps) energy utlty (Mb/J).5.5 ( M) (5.5 M) overall.96 staton to staton : M overall.4 overall.37 overall.3 ( M) (5.5 M) overall.4 overall.35 overall.57 overall.55 (b) Channel rates between statons: -A, M; -A, 5.5 M; -, M Fg. 7. The throughput and energy utlty of statons under dfferent s ( proxy and 3 clents) throughput (Mbps) energy utlty (Mb/J) ( M) ( M) overall.9 staton to staton : M overall.49 overall 3.7 overall.96 ( M) ( M) overall.5 overall.5 overall.49 overall.47 (a) Channel rates between statons: -A, M; -A, 5.5 M; -, M throughput (Mbps) energy utlty (Mb/J) ( M) (5.5 M) overall.6 staton to staton : M overall.77 overall.5 overall. ( M) (5.5 M) overall.7 overall.9 overall.36 overall.35 (b) Channel rates between statons: -A, M; -A, 5.5 M; -, M Fg. 8. The throughput and energy utlty of statons under dfferent s ( proxy and 5 clents) TBF wthout any throughput mprovement for the forwardng servce. Furthermore, wth our proposed SFW, the overall performance n the WLAN s also better than that of TBF- FW. These results ndcate that SFW not only provdes a strong ncentve for data forwardng, but also balances the tradeoff between the performance of ndvdual statons and the entre WLAN. Fgure 9(a) shows the growth of the proxy throughput gan n SFW (the proxy throughput of SFW over that of TBF) wth the ncreasng number of clents n the WLAN. In ths experment, the proxy (workng at Mbps channel rate wth the A) serves all other statons (workng at Mbps wth the A and Mbps wth the proxy) n the same WLAN. Wth channel tme compensaton, even n clent and proxy case, the proxy throughput can stll be mproved by 4% over TBF. Fgure 9(b) shows the proxy energy utlty gan n TBF-FW (the proxy energy utlty of TBF-FW over that of TBF) n the same crcumstances as above. The energy utlty gan of TBF-FW s less than, meanng the energy utlty s worse than that of TBF. Fgure 9(b) also ndcates that n TBF-FW, the proxy may have to consume more than % energy for ts clents, whch could prevent the proxy from provdng such servce. VIII. CONCLUSION In ths paper, we am to () address the throughput degradaton nduced by low channel rate statons n a WLAN, and () utlze the nevtable energy waste n channel lstenng durng a communcaton sesson. We characterze energy effcency as energy per bt, nstead of energy per second. Utlzng dle communcaton power, we present a data forwardng mechansm and an energy-aware token rewardng scheme to supplement the IEEE 8. protocols. In data forwardng, a hgh channel rate staton forwards data for a low channel rate staton, resultng n a sgnfcant mprovement of ts throughput. To gve hgh channel rate statons an ncentve to be proxes, we desgn an energy-aware token rewardng scheme, n whch low channel rate statons compensate for proxes wth addtonal tme slots. Thus, a proxy can also mprove ts own throughput wthout compromsng ts energy effcency. We have presented a mathematcal model to gude the protocol desgn, and have proposed algorthms for proxy selecton, channel allocaton and schedulng, and data forwardng n