1 Heerogeneos wireless nework managemen W. Qadeer*, T. Simnic*^, J.Ankcorn^, V. Krishnan^ and G. De Micheli* *Sanford Universiy, ^HP Labs Absrac Today s wireless neworks are highly heerogeneos wih diverse range and QoS. The mainenance of a wireless link by a mobile device reqires sppor of mliple nework inerfaces. Since he baery lifeime is limied, power managemen of heir inerfaces wiho any significan degradaion in performance has become essenial. In or research we developed an inegraed approach for he managemen of power and performance of mobile devices in heerogeneos wireless environmens. Or policy decides wha wireless nework inerface (WNIC) o employ for a given applicaion and how o opimize he WNIC sage. This decision is governed by he crren power and performance needs of he sysem. The policy dynamically es beween inerfaces dring program execion if daa commnicaion reqiremens and/or nework condiions change. For he verificaion of or power and performance managemen algorihm, we have experimenally characerized leooh and 802.11b wireless inerfaces. We implemened or policy on HP s IPAQ porable device ha is commnicaing wih HP s HoSpo server [14]. The applicaions we esed range from MPEG video o email. The resls show ha or policy offers a large improvemen in power savings as compared o singly sing 802.11b or leooh while enhancing performance. I. INTRODUCTION Mobile commnicaions oday has heerogeneos wireless neworks providing varying coverage and QoS. Varios commnicaion services are available. The infrasrcre enables mobile devices o rn applicaions wih diverse bandwidh and nework conneciviy reqiremens, sch as disribed speech recogniion, video sreaming, gaming ec. To saisfy he bandwidh and QoS consrains of he applicaions, he mobile devices need o allow seamless ing among varios wireless nework inerfaces.. Addiionally, he high commnicaion and compaion cos of applicaions is a brden on he baery life of porable devices. Capaciy of a baery has no increased remendosly. Improvemens of only a facor of 2-4 have been observed dring he pas 30 years. The ever-increasing need for baery lifeime in mobile devices demands a igher conrol over is energy consmpion. Alhogh low-power circi design forms he basis of power managemen in a mobile device, higher-level managemen of power dissipaion offers many more advanages. These echniqes allow seamless inegraion beween ser applicaions and power managemen policy design hs allowing energy consmpion o be redced while mainaining a desired QoS. Dring program execion commnicaion inerfaces are placed in low-power saes depending pon heir acces paerns and applicaion performance needs. Varios componens of he mobile sysem, sch as he ser, he wireless channel and he individal inerfaces can be modeled sing a sae based absracion. The echniqes developed o dae for he enhancemen of heerogeneos neworks concenrae on improving heir accessibiliy and QoS. These mehods enable mobile devices o commnicae wih each oher by inrodcing changes in he nework proocol sack. They also allow esablishmen and mainenance of connecions beween mobile hoss sing any available links o improve robsness and performance. However, none of he echniqes adeqaely addresses power managemen. Power redcion mehodologies presened in he pas largely focs on improving energy consmpion of one single device e.g. WaveLAN, CPU ec. Policies for he redcion of power dissipaion range from simple ime-o mehods o complex echniqes based pon sochasic models. This work presens a new mehodology for managing power and performance of mobile devices consising of heerogeneos WNICs. The policy formlaed decides wha nework inerface o employ on a porable device for a given paern of sage. The decision is governed by he crren power dissipaion and QoS reqiremens of he sysem. The maximm likelihood esimaor is employed for racking sysem changes. I deecs variaions in he average hroghp of available wireless inerfaces and he daa sage paerns. The policy for power and performance managemen (PPM) decides: a. Wha wireless nework inerface card o se b. Wha low-power sae o employ c. Transiion imes beween and low-power saes d. ffer size o se for good applicaion QoS We implemened he policy on HP s IPAQ porable device ha is commnicaing wih HP s HoSpo server [14] via leooh and 802.11b. The applicaions we esed range from MPEG video o email. Or resls show boh large savings in power when sing a single WNIC, as well as seamless ing wih concrren power savings among WNICs.
2 II. RELATED WORK Mobile devices reqire wireless commnicaion inerfaces o faciliae conneciviy wih Inerne and wih he oher devices. A mechanism is reqired for forwarding packes beween differen wireless neworks de o increasing device mobiliy. Mobile IP [1] provides one example of sch mechanism. Changes are inrodced in he nework and link layers of he nework proocol sack ha assis he hos s home nework in forwarding packes o is nework of residence. However, wih mobile IP even if commnicaing devices are in he same wireless nework, daa needs o raverse a mlihop pah. In order o perform localized commnicaion beween devices, which are one hop disance away, Conac Neworking [2] has been proposed. y allowing seamless ing beween mliple diverse inerfaces, his echniqe enhances robsness and QoS of he nework. Mobile hoss experience varying daa raes dring commnicaion in par de o lossy nare of he wireless link. In order o avoid disrpions, a disribed file sysem has been developed [3], [4]. I allows applicaion aware and applicaion independen adapaion o a emporary loss or degradaion of he wireless link hs enhancing robsness. A mehod for improving hand-offs proposes bffering daa on mliple base saions in close proximiy o he mobile hos [5] hs achieving seamless ing beween base saions. Telephony and daa services spanning diverse access neworks have been inegraed in [6]. However, he focs of hese echniqes has been on he enhancemen of performance and QoS of heerogeneos neworks. Power managemen of commnicaing hoss has been mainly overlooked. Several echniqes have been proposed o efficienly manage power dissipaion in porable devices. These mehods employ diverse mechanisms o predic periods of inaciviy dring commnicaion. ased pon hese predicions he mobile device is p ino a low-power sae. The mos basic power managemen policy is a ime-o. If he device remains idle for a cerain period, i is p ino a low power sae. Similarly, a device can ener low-power mode when idleness is being anicipaed in a connecion [7]. However, incorrec esimaes case performance and power penalies. In conras, sochasic models derive provably opimal power managemen policies. Pre Markov decision processes [8], [9] employ eiher discree or coninos ime memory-less disribions. However, discrepancies have been observed in prediced and acal power savings owing o hisory dependen nare of real world processes. Time-indexed semi Markov decision processes [10] are based pon hisory based disribions. This echniqe has demonsraed energy savings in real-world applicaions. The power managemen echniqes presened o dae mosly focs on he redcion of power dissipaion in one WNIC. This leads o inefficien power managemen for porables wih mliple diverse commnicaion inerfaces. Mehods being employed for he performance enhancemen of homogeneos neworks p a lo of emphasis on power managemen. IEEE 802.11 [11] sandard implemens power managemen by sending a raffic indicaion map (TIM) wih he beacon o he clien. I enables he clien o ener doze mode if no more daa is available. Since he device sill has o wake p afer every beacon inerval for TIM, a new echniqe proposes decopling of conrol and daa channels [12]. The conrol channel ses low-power radio and wakes p he device whenever daa is presen. Applicaion level informaion is sed for power managemen in [13]. In or work we developed an inegraed policy for power and performance managemen. Or power and performance managemen (PPM) algorihm dynamically selecs he appropriae wireless nework inerface wih he goal of minimizing he overall energy consmpion while meeing applicaion s QoS reqiremens. We presen measremen resls ha show large energy saving wih good QoS while sing leooh and 802.11b on HP s IPAQ for a ypical se of applicaions. The res of he paper is organized as follows. Secion 3 discsses he characerizaion of leooh and 802.11b inerfaces. Deails of he herisic policy for choosing among nework inerfaces are presened in Secion 4 whereas he resls and conclsion are discssed in Secions 5 and 6 respecively. A. leooh III. CHARACTERIZATION OF DEVICES leooh has been developed as a radio link wih a shor range o provide wireless conneciviy o porable and fixed devices. I operaes in 2.4GHz ISM band. leooh sppors poin-o-poin and poin-o-mli-poin connecions called picones. A picone can consis of wo o eigh leooh devices. One device is he maser and he res are is slaves. In addiion, a maser can sppor several oher in slaves, which have been parked. These slaves remain synchronized o he maser b do no become a par of he picone. A scaerne is composed of mliple picones wih an overlapping coverage area. leooh provides boh synchronos and asynchronos connecions. A synchronos connecion can sppor 64kb/s in each direcion whereas he asynchronos one can sppor a maximm of 723.2 kb/s asymmeric and 433.9kb/s symmeric daa raes [16]. leooh sppors mliple low-power saes: hold, sniff and park. As shown in he sae space figre below, hese saes can only be acivaed once a connecion exiss beween leooh devices.
3 TALE 1. LUETOOTH LOW-POWER MODE MEASUREMENTS TRANSMIT DATA (DH/DM PACKET) INQUIRE (UNKNOWN ADDRESS) PARK OFF STANDY CONNECTED Maser/Slave HOLD DEEP SLEEP FIGURE 1. LUETOOTH STATE SPACE PAGE (KNOWN ADDRESS) SNIFF a. Hold mode is employed o sop daa ransfer by he reqesed device for a negoiaed inerval. I is especially sefl if he reqesing device wans o perform inqiry, page and scan or brs mode ransfer operaion in scaernes. b. Sniff mode is sefl on low daa rae links sch as email where a qick response is reqired whenever daa is presen. Dring an aemp window he device looks for any incoming daa. If no daa is presen, i goes ino lowpower mode; however, if daa is presen, he device lisens o he maser for he specified ime-o period. Sniff mode can also be sefl in scaernes for devices ha are a par of wo picones. TRANSITION TIME (M SEC) AVG. POWER (W) HOLD MODE 0.061 Hold mode enry 1.68 0.068 Hold mode exi 11.62 0.216 PARK MODE 0.061 Park mode enry 2.16 0.077 Park mode exi 4.12 0.126 SNIFF MODE 0.061 Sniff mode enry 0.94 0.078 Sniff mode exi 7.36 0.194 leooh sppors mliple packe ypes for boh asynchronos and synchronos connecions. These packe ypes differ in daa payload size and error correcion algorihms. Maximm achievable hroghp for varios packe ypes is ablaed in Table 2. Or measremen resls come close o hroghp vales repored in Table 2. For insance, he maximm and he average hroghp nmbers were measred on CSR leooh for DH5 packes a 87k/s and 79k/s respecively. The hroghp increases wih an increase in he payload capaciy of he base-band packe. However, hroghp can significanly decreased in he presence of noisy channels if less error-correcion bis are presen. The range of leooh devices is enhanced by an increase in ransmission power; b ha cases a frher energy drain from he baery. TALE 2. MAXIMUM THROUGHPUT FOR LUETOOTH ACL CONNECTION Symmeric max rae (kb/s) Asymmeric max rae (kb/s) Forward Reverse DM1 108.8 108.8 108.8 DM3 258.1 387.2 54.4 DM5 477.8 477.8 36.3 DH1 172.8 172.8 172.8 DH3 390.4 585.6 86.4 DH5 433.9 723.2 57.6 c. Park mode is sed o enhance he nmber of simlaneos conneced slaves. As link se p akes abo 10s in leooh, i is bes o reain an esablished connecion. In his mode no daa ransfer akes place as he parked slave gives p is connecion ID b i remains synchronized o he maser. Transiion imes and average power dissipaion for ing beween and hold, sniff or park mode are shown in Table 1. The ransiion imes are very shor. Conseqenly, leooh devices can seamlessly shif beween and low-power saes wih very lile energy dissipaion. In addiion, he CSR leooh chips sppors deep sleep sae wih only 270W power consmpion [15]. This sae can only be enered when a CSP based device is in one of he low-power saes. Deep sleep sae can be enered only if here is no aciviy on UART for a leas 250ms.. WaveLan-802.11b 802.11b has been developed o provide fas wireless conneciviy o mobile devices. Theoreically, 802.11b can sppor a maximm daa rae of 11Mbps in 2.4GHz band. I is designed o work in adhoc as well as infrasrcre nework opologies. Today a large majoriy of all WLAN commnicaion happens in infrasrcre mode, hs his is he mode we will be focsing on in his work. An access poin acs as a bridge beween wired and wireless neworks. An associaion is developed beween he access poin and he 802.11b card before commencing daa commnicaion. In order o faciliae mobiliy, access-poins also sppor roaming. WLAN has wo saes, ransmi and receive, in addiion o wo low-power modes, doze and off. Table 3 shows average power dissipaion measremens for he above
4 menioned power saes. According o he 802.11b sandard, a synchronizaion beacon is ransmied o he awake card by a cenral access poin (AP) every 100ms. The beacon is followed by a raffic indicaion map (TIM) indicaing any reqired daa ransfers. Doze mode is acivaed nil he nex beacon if no daa ransfer is reqired. This power managemen (PM) policy does no always give opimal power savings de o he following facors: a. An increasing nmber of cliens cases radios o say on longer since here is more conenion de o mliple simlaneos synchronizaion aemps by he mobiles. b. 802.11b s response ime o he AP sffers de o he delays imposed by he doze mode. c. Even wiho any rnning applicaions, 802.11b spends a considerable amon of ime lisening wih an increase in broadcas raffic and is hs nable o ener doze mode. Doze mode can only be acivaed by he hardware. However, ransiions o he off sae from eiher he or he doze sae can be conrolled a he OS level. Transiion imes and average power dissipaion for ing beween and low-power saes have been ablaed in Table 3 for Cisco 350 WNIC. TALE 3. 802.11 LOW-POWER MODE MEASUREMENTS TRANSITION TIME (MS) AVG. POWER (W) OFF TO DOZE STATE Doze sae enry 0.1 1.4 Doze sae exi 1 1.6 ACTIVE TO OFF STATE Off sae enry 1 1.7 Off sae exi 300 2.3 IV. POWER AND PERFORMANCE MANAGEMENT The goal of PPM is o enhance QoS while minimizing power dissipaion in a porable device. PPM s primary ask is o deermine wha nework inerface is mos siable for he applicaion needs and how o manage is power and performance saes. When an applicaion sars on a porable device, PPM pre-selecs hose WNICs for daa commnicaion whose average hroghp is greaer han he daa consmpion rae of he applicaion. This ensres ha he QoS reqiremens of he applicaion are saisfied. In sreaming applicaions a special emphasis is placed pon he daa bffer size. I no only deermines he average sleep ime of he commnicaion device b also he energy dissipaed in he RAM. Since he size of he bffer is deermined by he difference beween he hroghp and he daa consmpion rae, all he pre-seleced WNIC are frher examined o deermine no only heir commnicaion power dissipaion b also he resling RAM power consmpion. The one ha offers maximm power savings is seleced. Addiionally, dring he examinaion of he commnicaion energy only hose low-power saes, which are mos siable for he crren scenario, are considered. PPM also defines he ing ime beween and sleep saes for he seleced WNIC. PPM dynamically keeps rack of he variaions in he applicaion daa consmpion rae and he hroghp of wireless inerfaces sing he log of he maximm likelihood esimaor as shown in eqaion (1). A change in rae is defined o occr a poin c when comped likelihood over he las w daa poins is greaer han a prese hreshold. In or work we se 99.5% as a hreshold. The change is observed beween he old, λ old, and he new rae, λ new. Deails of his algorihm are frher discssed in [17]. m λnew ln( P max ) = ( w c + 1)ln ( λnew λold ) j (1) λ old j= k Whenever a change occrs, he PPM evalaes which sbse of WNICs cold handle he applicaions crrenly rnning by insring ha he available WNIC s hroghp rae, λ, is greaer han he applicaion s daa consmpion rae, λ. λ λ (2) Nework inerfaces saisfying eqaion (2) are frher analyzed o idenify he inerface ha offers maximm power savings for he given applicaion while keeping he reqired qaliy of service. The oal energy, E oal, consmed dring a given session along wih he average power dissipaion, P avg, is given by he following eqaions: E = E + E + P T (3) oal avg comm oal RAM (( ) T ) P = E λ + (4) Where E comm and E DRAM denoe he energy consmed by he WNIC and he DRAM respecively dring he commnicaion period, specifies he size of he bffer for sreaming applicaions ha is ly read and wrien o in seady sae. P and T indicae he average power dissipaed and he ime aken when ing from one WNIC o anoher. Only WNICs wih hroghp high enogh o mee applicaions demands are considered. Figre 2 depics he process of ing inerfaces. The deails of his procedre are frher elaboraed in [2] and [18].
5 FIGURE 2. PROCEDURE FOR SWITCHING WIRELESS NETWORK INTERFACES In sreaming applicaions he size of he bffer direcly affecs E comm and E DRAM. If he size of he bffer ndergoes an increase, he average power dissipaion of he commnicaion device diminishes de o longer sleep periods and hs less overhead in ransiion beween power saes. On he oher hand, he RAM energy increases wih increasing bffer sizes as he nmber of banks increases. Ths he deerminaion of he bffer size is of principal imporance for enhancing power savings. FIGURE 3. UFFER LAYOUT AND THE ASSOCIATED TIME INTERVALS The size of he bffer is chosen in a way sch ha he ransmissions cold be schedled in brss. In beween brss he WNIC can ransiion ino low power mode, hs saving energy and freeing bandwidh from conenion. The size of he bffer is deermined according o he eqaions given below. = + + (5) sleep csh = + (6) sleep T sleep on rans + ( λ λ ) ( λ λ ) off rans = (7) ( T λ λ ) ( λ λ ) off rans = off rans (8) ( T λ λ ) ( λ λ ) on rans = on rans (9) Where: a. T sleep is he average sleep inerval of he commnicaion device. b. T accons for he wors case delay enconered in dynamically ing beween wo nework inerfaces. c. T csh provides a cshion for any small variaions presen in he sysem. d. χ and δ denoe small variaions in hroghp and daa consmpion rae respecively. e. T off-rans and T on-rans are he ransiion imes beween and low-power and low-power and saes respecively. f. T be is he break-even ime and is defined in erms of power consmed dring he ransiion, P rans = P on-rans + P off-rans, he power consmed in he and sleep saes, P and P sleep. Prans P Tbe = Trans + Trans (13) P P sleep g. max is he maximm amon of memory available. When an applicaion sars, i wais for T iniial-delay before beginning o read from he bffer. This ime inerval is inflenced by he maximm delay a ser can olerae a sarp. Ths he ime o ener seady sae, T seadysae, is given by eqaion 14. Dring his inerval he commnicaion device says in he mode. T ( T λ ) ( λ λ ) seadysa e = Tiniial delay + iniial delay (14) The oal energy consmed by he commnicaion device, E comm, dring he bffer refill period is given by eqaion (15). Noe ha for simpliciy he ransiion power and ime have been combined ino one variable, P rans and T rans. The commnicaion energy needs o be balanced by he energy consmed by memory, as larger bffer sizes case higher energy consmpion. comm ( T+ Tcshion ) + Pransrans T Psleep Tsleep E = P + (15) The amon of energy consmed by memory, E RAM, is deermined from he energy consmed by he banks ha are ly paricipaing in reading and wriing of daa, E, and he energy of non- banks, E non- : E RAM = E + E (16) non = λ (10) T E wrie ( P + N P )( λ ) = P T + (17) csh = T csh ( λ ( 1 χ ) λ ( 1+ δ )) (11) Enon ( Nbanks Nabanks) Pnon read abanks refresh = (18) ( Tbe λ ) ( λ λ ) max λ (12) abanks Size bank N = (19)
6 Where: a. P wrie and P read specify he average power dissipaed when he RAM is wrien and read respecively. b. N banks specifies he oal nmber of available memory banks. c. P non- is he average power consmed by he memory banks ha are non-. d. P refresh is he average power spen in refreshing he banks no paricipaing in read and wrie operaions. e. N abanks is he nmber of memory banks where each has size Size bank V. RESULTS The server sed is a research prooype of HP s HoSpo [14] server whereas he policy has been implemened on IPAQ 3970 ha sppors boh 802.11b (CISCO Airone 350 PCMCIA WLAN) and leooh (CSR) inerfaces. The operaing sysem rnning on he IPAQ is Linx. The power measremens have been performed wih a DAQ card a 10ksamples/sec. In or experimens, we have sed ransmission conrol proocol (TCP) for all daa commnicaion. For leooh his has been done sing bnep. Figre 4 depics he decrease in commnicaion energy as he bffer size increases whereas he increase in RAM energy wih an increase in he bffer size is shown in Figre 5. 0.125 0.12 0.115 0.11 0.105 0.1 0.095 0.09 0.085 100 200 300 400 500 600 700 800 900 FIGURE 6. APPLICATION DATA CONSUMPTION RATE (k/s) 250 200 150 FIGURE 4. COMMUNICATION POWER VS UFFER SIZE A firs we consider individal applicaions and employ or PPM o deermine he appropriae WNIC for each one based pon is daa consmpion rae and he average hroghp sppored by he WNIC. For his experimen we have assmed ha he hroghp and he daa sage paern do no change significanly dring program execion. 100 50 0 0 1 2 3 4 5 6 7 8 9 FIGURE 5. DRAM POWER VS UFFER SIZE Ths PPM is employed o enhance performance while minimizing power dissipaion. PPM pre-selecs WNICs for a pariclar applicaion based pon heir average hroghps and he daa consmpion rae of he applicaion. The WNIC ha offers minimm power dissipaion wih regards o commnicaion and RAM energies is seleced. The appropriae low-power sae of he WNIC along wih he ing poins is also defined by PPM. Addiionally, i can dynamically he seleced WNIC if a change in is hroghp and/or he average daa consmpion rae of he applicaion is deeced. In he nex secion we presen he resls obained by sing or PPM wih a ypical se of applicaions having diverse daa sage paerns. x 10 7 For MP3 adio sreaming leooh and 802.11b offer similar performance, b 802.11b gives more power savings when sed in he off mode. es power savings in leooh are obained in park wih deep sleep enabled. In he off sae 802.11b does no consme any power. In conras leooh canno be rned off as he connecion drops whose reesablishmen reqires 1-10 sec. The resls are shown in Figre 7. FIGURE 7. POWER CONSUMPTION OF WIRELESS INTERFACES FOR MP3 AUDIO leooh is more siable han 802.11b for email raffic as i offers larger power savings. The appropriae lowpower sae for leooh is again Park wih deep sleep
7 enabled. Similarly for 802.11b, he appropriae low-power mode is off mode. However, 802.11b incrs a large power dissipaion penaly when ing from off o sae. Ths leooh wih park and deep sleep behaves beer han 802.11b wih off mode enabled. The power measremens are shown in Figre 8. and non-exisen power dissipaion in he off sae. The resls are shown in Figre 11. FIGURE 11. POWER CONSUMPTION FOR MPEG4 VIDEO (160X120) FIGURE 8. POWER CONSUMPTION OF WIRELESS INTERFACES FOR EMAIL leooh seems o be he connecion of choice for elne and WWW based applicaions owing o is faser response ime and low-power dissipaion as compared o 802.11b. Again he siable low-power sae for leooh is Park wih deep sleep enabled. However, de o a significan decrease in he response ime for 802.11b wih off mode, he low-power sae of choice for 802.11b is PM. The ing overhead associaed wih he resls is shown in figres 9 and 10. For sreaming applicaions, power dissipaed by he commnicaion device is an increasing fncion of he bffer size. However, by sing PPM in or pariclar sep, we fond ha he bffer size is limied by he power dissipaed in he DRAM. If he size of he bffer is increased by one, he cmlaive average power dissipaion of he wireless card and he DRAM dobles. FIGURE 12. THROUGHPUT CHANGE FORCES A WNIC SWITCH FIGURE 9. POWER CONSUMPTION OF WIRELESS INTERFACES FOR TELNET FIGURE 10. POWER CONSUMPTION OF WIRELESS INTERFACES FOR WWW For MPEG4 video sreaming, leooh can only be employed for small images de o is lower average hroghp. The low-power mode for leooh is again park wih deep sleep enabled. However, 802.11b wih off mode offers more power savings de o higher hroghp Nex we presen resls of he experimen ha inclde a dynamic beween wireless inerfaces dring program execion. In his experimen he daa consmpion rae of he applicaion says consan, b he hroghp of he seleced WNIC ndergoes a change. Le s sppose ha a person wih a porable device is sreaming MP3 adio sing leooh. Afer a cerain period of ime he moves away from he server and he hroghp of leooh experiences a sharp decrease. The esimaor deecs he change in hroghp and forces PPM o reevalae he siabiliy of he seleced nework inerface. The reevalaion sggess a change in he nework inerface and 802.11b is chosen over leooh. The comparison of power dissipaion is shown in Figre 12. Lasly, we analyze he performance of PPM when applicaion daa consmpion rae changes. We creaed an applicaion race consising of MP3 adio, Email, elne, WWW and MPEG4 video. The race is execed firs by sing 802.11b only wih PM enabled, secondly by employing leooh only wih park as he low-power mode and in he end by PPM wih leooh and 802.11b as he available
8 wireless inerfaces. We fond ha PPM offers a facor of 2.9 imes improvemen in power savings over js employing leooh wih park mode whereas in comparison o 802.11b wih PM he power savings are a facor of 3.2 imes higher. Moreover, PPM enhances he QoS as wireless inerfaces are ed o mach he daa sage paern of he applicaion. The resls are shown in Figre 13. FIGURE 13. CHANGE IN DATA CONSUMPTION RATE FORCES A WNIC SWITCH When he applicaion race is execed sing 802.11b only, he wireless inerface is placed in he doze mode whenever TIM indicaes periods of inaciviy. Similarly for leooh, he inerface is placed in he park mode whenever no daa commnicaion is needed. However, de o he varying daa sage paerns, hroghps and low-power modes of he wireless cards, power savings vary considerably from 802.11b o leooh. PPM selecs he appropriae wireless inerface card a he sar of each applicaion. For his experimen he inerfaces are no dynamically ed dring program execion. VI. CONCLUSION This work presens a new mehodology for enhancing QoS while maximizing power savings in heerogeneos wireless sysems. A policy for selecing he mos appropriae nework inerface for a pariclar applicaion has been developed. We have esed or policy on IPAQ 3970 spporing 802.11b and leooh wireless inerfaces sing varios ypical applicaions. We have shown ha or PPM offers 2.9 and 3.2 power savings over solely sing leooh and 802.11b respecively when rnning a sring of applicaions inclding MPEG4 video and MP3 adio. [5] Eric A. rewer, Randy H. Kaz, Elan Amir, Hari alakrishnan, Yain Chawahe, Armando Fox, Seven D. Gribble, Todd Hodes, Giao Ngyen, Venkaa N. Padmanabhan, Mark Semm, Srinivasan Seshan, and Tom Henderson, A nework archiecre for heerogeneos mobile comping, IEEE Personal Commnicaions Magazine, 5(5):8 -- 24, 1998.. [6] Helen J. Wang, haskaran Raman, Chen-nee Chah, Rahl iswas, Ramakrishna Gmmadi, arbara Hohl, Xia Hong, Emre Kiciman, Zhoqing Mao, Jimmy S. Shih, Lakshminarayanan Sbramanian, en Y. Zhao, Anhony D. Joseph, and Randy H. Kaz, ICEERG: An Inerne-core Nework Archiecre for Inegraed Commnicaions, IEEE Personal Commnicaions: Special Isse on IPbased Mobile Telecommnicaion Neworks, 2000. [7] C. H. Hwang and A. W, A Predicive Sysem Shdown Mehod for Energy Saving of Even-Driven Compaion, in Inernaional Conference on Comper Aided Design, pp. 28 32, 1997. [8] E. Chng, L. enini and G. De Micheli, Dynamic Power Managemen for non-saionary service reqess, Design, Aomaion and Tes in Erope, pp. 77 81, 1999. [9] Q. Qi and M. Pedram, Dynamic power managemen of Complex Sysems Using Generalized Sochasic Peri Nes, Design Aomaion Conference, pp. 352 356, 2000. [10] T. Simnic, L. enini, P. Glynn, G. De Micheli, Even-Driven Power Managemen, IEEE Transacions on Comper-Aided Design, Jly 2001. [11] The Ediors of IEEE 802.11, IEEE P802.11D5.0 Draf Sandard for Wireless LAN, Jly, 1996. [12] E. Shih, P. ahl, M. J. Sinclair, Wake on Wireless: An Even Driven Energy Saving Sraegy for aery Operaed Devices, in Proc. of ACM MobiCom, 2002. [13] A. Acqaviva1, T. Simnic, V. Deolalikar, and S. Roy, Remoe power conrol of wireless nework inerfaces, in Proc PATMOS, 2003. [14] D. Das, G. Manjnah, V. Krishnan, P. Reddy, HoSpo! a service delivery environmen for Nomadic Users Sysem Archiecre, Mobile Sysems and Sorage Lab., HP Laboraories Palo Alo, CA Rep. HPL- 2002-134, 2002. [15] CSR, lecore Power Saving Modes, Jan, 2003. [16] leooh Special Ineres Grop, "Specificaion of he leooh Sysem 1.1, Volme 1: Core," hp://www.bleooh.com, Feb, 2001. [17] T. Simnic, L. enini, A. Acqaviva, P. Glynn, and G. De Micheli, Dynamic volage scaling for porable sysems, in Proceedings of he 38h Design Aomaion Conference, Jne 2001. [18] J. Torrilhes and C. Carer, P-Handoff: A proocol for fine-grained peer-o-peer verical handoff, Proc. of PIMRC, 2002. VII. REFERENCES [1] D.. Johnson, and D. A. Malz, Proocols for adapive wireless and mobile neworking, IEEE Personal Commnicaions, 1996. [2] C. Carer, R. Kraves, and J. Torrilhes, Conac Neworking: A Localized Mobiliy Sysem, in Proc. of he Firs Inernaional Conference on Mobile Sysems, Applicaions, and Services, 2003. [3] J. Flinn, M. Sayanarayanan, Energy-aware adapaion for mobile applicaions, Proceedings of he 17h ACM Symposim on Operaing Sysems Principles, 1999. [4]. Noble, M. Sayanarayanan, D. Narayanan, J. E. Tilon, J. Flinn, K. Walker, Agile Applicaion-Aware Adapaion for Mobiliy, in Proc. of 16h ACM Symposim on Operaing Sysem Principles, 1997.