Energy-Aware Wireless Video Streaming

Transcription

1 Energy-Aware Wreless Vdeo treamng Abstract Ths paper presents a dynamc energy management polcy for a wreless vdeo streamng system, consstng of battery-powered clent and server. The paper starts from the observaton that the vdeo qualty n wreless streamng s a functon of three factors: encodng apttude of the server, decodng apttude of the clent, and the wreless channel. Based on ths observaton, the energy consumpton of a wreless vdeo streamng system s modeled and analyzed. Usng the proposed model, the optmal energy assgnment to each vdeo frame s done such that the maxmum system lfetme s acheved whle satsfyng a gven mnmum vdeo qualty requrement. Expermental results show that the proposed polcy ncreases the system lfetme by %. Introducton Wth the avalablty of moble, communcaton and computng systems, we have seen an explosve growth n wreless multmeda applcatons, e.g., streamng audo and vdeo. Ths trend n turn poses two challenges: () establshng and mantanng a stable channel for real-tme operaton and () power-aware operaton so as to ncrease the lfetme of the battery-powered wreless system whle meetng a mnmum qualty of servce (Qo) requrement. Furthermore, t s desrable to provde a mechansm for graceful degradaton n Qo so that a dynamc power manager (DPM) can ncrementally trade off Qo for hgher energy effcency. Fne Granularty calablty (FG) codng technque [], whch was adopted as the standard n MPEG- 4, provdes an effectve mechansm for graceful vdeo qualty degradaton based on ts herarchcal layer structure, whch conssts of a base layer and one or more (optonal) enhancement layers. Although extensve studes have been conducted on the herarchcal layer structure of MPEG-4 and ts error reslency under fluctuatons n the channel bandwdth [][3][4], energy effcency n a battery-powered server-clent system has receved lttle attenton. For vdeo streamng applcaton, there are two sources of energy consumpton n wreless moble hosts: computaton energy for processng a vdeo stream and communcaton energy for transmttng and recevng data. The computaton energy of a server and a clent s usually a strong functon of the PU frequency, whch can be changed by employng methods such as dynamc voltage and frequency scalng (DVF). The communcaton energy, on the other hand strongly affects the bt-error-rate (), and hence the vdeo qualty. There are detaled studes of the trade-off between energy consumpton and n the communcatons feld [5]. These studes can be dvded nto two man categores. Ths research was supported n part by DARPA PA/ program under contract DAAB7---P3 and by NF under grant no Al Iranl, Khwan ho, and Massoud Pedram Dept. of EE-ystems, Unv. of outhern alforna, Los Angeles, A 989 Emal: {ranl, khwanch, pedram}@usc.edu The frst category of technques, whch focus on the passband transcever, explots the fact that dfferent modulaton schemes result n dfferent vs. sgnal-to-nose rato (NR) characterstcs. The basc dea s that by adaptvely changng the modulaton and/or equalzaton, whle keepng constant the receved NR at the recever, one can acheve dfferent. The second category of technques, whch focus on the base-band transcever, studes the nteracton between code performance and encoder/decoder desgn complexty. The man dea s to add a number of error controllng bts to the orgnal data bts to protect them from channel changes. The key trade off s between the complexty of the encodng/decodng algorthms and the. The achevable vdeo qualty n the streamng vdeo systems s determned by followng three factors: encodng capablty of the server, decodng capablty of the clent, and the wreless channel error rate. It s well known that channel bandwdth fluctuaton due to varous factors result n the severe degradaton n the vdeo qualty. Ths s due to the streamng nature of ths real-tme operaton and the extra tme, whch s requred for retransmssons f errors occur n the data packets. The encodng (decodng) apttude of the server (clent) s defned as the amount of data that can be processed n a gven deadlne. Ths apttude s proportonal to the nverse of the vdeo frame rate. When the server (or/and the clent) changes ts operatng frequency and voltage to extend ts lfetme the encodng (decodng) apttude s also affected, so s the qualty of the streamng vdeo. Ths scenaro s not unusual because many of the state-of-the-art processors that are desgned for moble applcaton are equpped wth DVF for low-power operaton [6]. In [7] a low energy MPEG-4 streamng polcy usng a clent-feedback method was proposed where the clent decodng capablty at each tme slot s sent to the server and the server adjusts ts sendng rate based on the feedback value from the clent. By usng ths feedback approach, a sgnfcant amount of communcaton energy savng was acheved. However, the authors consdered only energy consumpton of the clent sde, not ncludng that of the server. In [], the authors proposed an energy-optmzed mage transmsson system for ndoor wreless applcatons, whch explots the varatons n the mage data and the wreless mult-path channel by employng dynamc algorthm transformatons and jont source-channel codng. A detaled energy model for the clent-server system was proposed and a global optmzaton problem solved by usng feasble drecton methods that resulted n an average of 6% energy savng for dfferent channel condtons.

2 In ths paper, we propose an adaptve polcy for a wreless vdeo streamng system n whch the optmal energy assgnment to each vdeo frame consderng both the server and clent s employed such that the system consumes the mnmum energy, whle meetng the requred vdeo qualty constrant. Herarchcal game theory s used to solve the correspondng mathematcal optmzaton problem. Expermental results show an average of % ncrease n the overall system lfetme. The remander of ths paper s organzed as follows. ecton ncludes backgrounds on MPEG-4 FG, model for energy consumpton of the server and the clent n the streamng system. ecton 3 descrbes our energy assgnment problem, and secton 4 dscusses the game theoretc formulaton for ths problem. Expermental results are descrbed n ecton 5 and t s followed by concluson n ecton 6. Background. Fne Granularty calablty (FG) To adapt to a tme-varyng channel capacty (whch s n turn due to changes n the channel condton for example because of congeston or fadng), a number of scalable vdeo codng technques have been proposed. Typcal technques nclude NR scalablty, temporal scalablty, and spatal scalablty n MPEG- and MPEG-4. In these layered scalable codng technques, the total encoded bt-streams consst of a base layer and several enhancement layers. The bt-rate of the base layer s determned by the mnmum channel bandwdth and s suffcent to ensure a mnmum achevable vdeo qualty. The enhancement layers provde hgher vdeo qualty when the channel has extra bandwdth for the transmsson of extra layers. The FG vdeo codng technque, whch was adopted as the standard n MPEG-4, provdes a very smooth varaton n the vdeo qualty compared to other scalable codng technque because any number of bts n the enhancement layers may be truncated accordng to the channel condton. Therefore, the Vdeo qualty (VQ) can be represented as a lnear equaton n terms of number of bts transmtted: where k s a regresson coeffcent, R send s the total bt-rate (bts/sec), R b s the base layer bt-rate, and R e s the enhancement layer bt-rate. Note that R b must be less than the mnmum achevable bandwdth n the wreless channel, otherwse, no useful vdeo transmsson s possble and VQ goes to zero. R e s vared n response to the channel condtons. R send s thereby set to provde the mnmum acceptable vdeo qualty by transferrng the mnmum amount of vdeo data to the clent subject to the exstng channel condtons and the remanng battery lfetmes of the vdeo server and/or the clent. Recall that the hgher the bandwdth of the channel s, the hgher the VQ s for a fxed level of total energy consumpton.. Energy Model of the erver The energy consumpton of the server for processng and transmttng a vdeo frame may be wrtten as: E = E + E omp omm where E omp and E omm denote the per-frame energy consumpton costs of the computaton and communcaton processes n the server. E omp and E omm are n turn calculated as follows: E = V f T omp eff ( ) E = P + P + P T omm Enc Mod Amp where eff denotes the effectve swtched capactance per clock cycle tme n the server, V s the supply voltage level (assumng full swng transtons) of the server PU, f s the clock frequency of the server PU, and T s the tme duraton of a frame (.e., nverse of the frame rate). P Enc, P Mod, and P Amp denote power consumptons of the correspondng blocks n the transmtter. The term representng the power consumpton of the amplfer, P Amp, s qute mportant. The other terms tend to be smaller n magntude and depend lnearly on the symbol rate wth an addtonal constant. Hence, for our optmzaton purposes, the communcaton energy consumpton of the server may be approxmated as: ( ) () (3) E = P R + P + P T (4) omm Tx s const Amp where P Tx and P const are the symbol-rate-dependent and constant power consumpton components of the base-band transmtter. R s denotes the symbol rate. To characterze the bt error rate () n terms of the power consumpton of the transmtter, the relatonshp between the receved sgnal-to-nose rato (NR) and the of the pass-band transcever,.e., the modulatng/demodulatng par, can be used. For example, consder a Quadrature Ampltude Modulaton (QAM) scheme where the s related to the receved NR by the followng equaton [5]: = ( P ) M NR VQ = k. Rsend = k.( Rb + Re ) () P =.. Q 3 rcvd M M M where M s the number of constellaton ponts n the QAM modulaton, typcally M = b where b s the number of nformaton bts represented by each constellaton pont. NR rcvd s the receved sgnal-to-nose-rato at the recever. Let N, β, and R s denote the nose spectral densty, the spectral shapng factor, and the symbol rate, respectvely. The receved NR s related to the transmt power level P Amp, nose n the channel P Nose, and the path loss parameter, σ, by [5]: PAmp PAmp NRrcvd = P σ = Nose N β R s σ (5) (6)

3 For a gven and modulaton scheme,.e., for fxed b, one can calculate the requred NR, from equaton (5), and then use equaton (6) to fnd the mnmum requred transmt power level. The overall energy consumpton of the transmtter for transmttng a sngle symbol s then calculated from equaton (4)..3 Energy Model of the lent The energy consumpton of the clent for recevng and processng a vdeo frame may be wrtten as: E = E + E omp omm where E omp and E omm denote the per-frame energy consumpton costs of the computaton and communcaton processes n the clent. They are calculated as follows: omp eff (7) E = V f T (8) where eff denotes the effectve swtched capactance per clock cycle tme n the clent, V s the supply voltage level (assumng full swng transtons) of the clent PU, and f s the clock frequency of the clent PU. E omm s due to energy consumptons of the low nose amplfer, the demodulatng block, and the channel decodng block and may be wrtten as: E omm = ( PLNA + PDemod + PDec ) T (9) where P LNA, P Demod, and P Dec denote the power consumptons of the correspondng blocks n the recever. onsderng that all other blocks except the channel decoder are fxed and do not respond to changes n channel condtons, for optmzaton purposes, the clent energy consumpton may be approxmated as: ( ) E P R + P + P T () omm Rx s const Dec where P Rx and P const are the symbol-rate-dependent and constant components of power consumpton of the pass-band recever. Typcally, a channel decoder s a mult-stage mplementaton of a recursve decodng functon. Therefore, the accuracy of decodng s ncreased as the number of decodng stages (teratons) ncreases. On the other hand, ncreasng the number of stages would ncrease the power consumpton of the decoder. In ths work, a Vterb decoder s studed as the channel decoder. In adaptve Vterb algorthms (AVA), developed n [3]-[5], the decodng performance s ncreased by reducng the number of operatons requred to decode a sngle bt. Ths s acheved by reducng Truncaton Length (TL) or by reducng the number of urvvor Paths (P),.e., those paths that are kept n order to fnd the optmum path. There are two man varatons of the AVA. In the frst varaton, whch s called the T-Algorthm [6], a fxed Threshold T, s chosen and then those paths that have path metrcs equal to or less than T are ncluded n the P memory. In the second varaton, called the M-Algorthm [6], a fxed number (M) of paths are kept and all other paths are dscarded. These paths are selected by choosng the frst M paths wth the mnmum path metrc values. onsder an adaptve Vterb decoder wth the functonal block dagram depcted n Fgure a. The decoder can be dvded nto three basc unts. The nput data (.e., the nosy observaton of the encoded nformaton bts) s used n the Branch Metrc Unt (BMU) to calculate the set of branch metrcs λ j,k. These are then fed to the Add-ompare-elect Unt (AU) to update the path metrc cost accordng to the followng recursve equaton: ( ) γk, + = mn γ jk, + λ jk,, γlk, +λ lk, () where γ,k s the path metrc cost for state s n tme step k, and λ j,k s the branch metrc cost between states s and s j from tme nstances k and k+, respectvely (cf. Fgure b). The urvvor Memory Unt (MU) processes the decsons that are beng made n the AU n order to carry out the Arecurson and outputs the estmated path, wth a latency of at least TL. Power consumpton for an adaptve Vterb decoder may be macro-modeled by summng up the power consumpton of each block tmes the number of paths that block s beng used. Ths would result n followng proposed power macromodel: ( K.(. )) P = P + P + TL P () Dec BMU AU MU where P BMU, P AU, and P MU are the per-operaton power consumptons of the correspondng modules and K represents the memory depth of the correspondng convolutonal encoder. Notce that the AU module performs two addtons and one comparson operaton n each step (cf. equaton ). 3 Energy Optmzaton Problem The encodng/decodng apttude of an mage processng core s a strong functon of ts operatng frequency and voltage level. Thus, one can characterze the vdeo qualty VQ of frame as: Input BMU AU Latch a. Block dagram of the Vterb decoder b. Fndng the optmum path Fgure : Adaptve Vterb decoder MU Output

4 s c VQ f ( E, E, h ) = (3) where E and E denote the server and the clent energy consumptons for frame whle ω denotes the wreless channel condtons for transmsson of frame. We consder a wreless system operatng over a fadng channel. Tme s assumed to be dscrete. In each tmeslot, whch s equal to the nverse of the gven frame rate, the channel state changes among dfferent states chosen from fnte set = { accordng to a known,, L, n } probablstc model [9]. The server and the clent are assumed to be battery-powered, each wth a fxed number of energy unts avalable for use. Each channel state determnes the throughput that can be acheved per unt energy expended by the server/clent. The vdeo encodng/decodng processng cores can operate wth frequences f and f n a range bounded by a lower bound f mn and an upper bound f max. Let EL max denote the maxmum number of enhancement layers that can be processed by the server and the clent whle operatng at f max. In each tmeslot, one can calculate the Pareto-optmal curve for energy consumpton of the server ( E ) and the clent ( ) n order to produce a gven vdeo qualty (VQ) under a E gven channel condton, ω. More precsely, ths curve denotes the trade-off between energy consumpton of server and that of the clent. Let s denote ths Pareto-optmal curve by g ( E, E ) =. Gven the remanng energy levels of the server ( E the clent ( E Tn Tn E E and E E = = ) and ), the objectve s to fnd a non-domnated energy allocaton par ( and ) for each tmeslot so as E to maxmze the overall system lfetme Tn, subject to: I) II) avg( VQ ) VQ III) ( ) E mn g E, E = (4) where avg(vq ) denotes the average vdeo qualty over the system lfetme. onstrant I corresponds to the total energy bound, whle constrant II guarantees the average vdeo qualty of the system. The thrd constrant ensures that Pareto-optmal ponts are chosen as the operatng ponts n each tmeslot. In the followng subsectons, we frst study the lfetme maxmzaton under the condton that the channel state s known a pror and s constant. Next, we assume that channel states are random wth a known probablty dstrbuton functon h(ω ) dentcally and ndependently dstrbuted (..d.) over tme, and that ω s not revealed untl just before the start of the tmeslot. We wll develop an approxmate algorthm for ths case, whch produces a nearly optmal soluton. 3. Known channel condtons Let s start by examnng the lfetme maxmzaton problem n the smple case where the channel condton s a pror known and remans constant. In other words, ω s a known and fxed state for all tmeslots. Although knowng the channel state for all tmes s an unrealstc assumpton, the soluton to ths problem provdes nsght whch s helpful for solvng the more realstc problem scenaro when the channel state s unknown and changng over tme. To solve ths problem, one should frst fnd the Paretooptmal curve correspondng to the requred vdeo qualty ( VQ ) and then fnd an operatng pont on that curve such mn that the lfetme of the system s maxmzed. Fgure shows typcal behavor of functon g n equaton (4). Obvously, ths functon s non-ncreasng n terms of energy consumptons E and E. The system lfetme s maxmzed when the server and the clent run out of energy at the same tme. To fnd an operatng pont at tmeslot, whch guarantees the maxmal system lfetme, one can set the rato of energy consumpton rates of the server and the clent equal to the rato of remanng energy levels of the server and the clent. Ths soluton s llustrated graphcally n Fgure. By drawng a lne wth slope equal to the rato of remanng energes of the clent and the server, and ntersectng ths lne wth the Pareto-optmal curve, we can fnd the optmal operatng pont and the correspondng energy consumptons of the server and clent, E and E. opt opt 3. Unknown channel condtons Let s examne the problem of lfetme maxmzaton under the assumpton that the actual channel condton ω s not known untl just before transmsson at tme. Moreover, let s consder a case where ω s..d. wth a known dstrbuton functon, h(ω ). To solve ths problem, we propose a dynamc polcy, whch attempts to mnmze the dfference between the remanng battery lfetmes of the server and the clent durng each tmeslot. E opt E E E E opt g(e,e )= E Fgure : Pareto-optmal energy curve of the clent-server par

5 A smple way to calculate the estmated remanng battery lfetme of a moble node (server or clent) at the begnnng of tmeslot s to dvde the remanng energy level of the node by ts energy depleton rate. A key challenge s to accurately estmate the expected depleton rate. Notce that smply usng the energy depleton rate of the prevous tmeslot, -, s not sutable because t does not account for the long-term behavor of the node and may thus result n erroneous estmates. The approach we have taken s to calculate the (hstory-based) aggregate energy depleton rate of the node as a movng exponentally-weghted average so that the recent past has more nfluence, but the dstant past s not completely gnored. At the begnnng of each tmeslot of duraton T, the server chooses an energy consumpton rate, E / T, for tself based on the current estmates of the remanng battery lfetmes of tself and the clent and the predcted channel state for tmeslot. In our approach, the channel state for tmeslot s taken to be the same as the channel state -. From the chosen energy consumpton rate, the server wll then put nto effect the respectve encodng (computaton) and transmt (communcaton) parameters by lookng up these values from a pre-computed and locally-stored polcy parameter table. nce the actual channel state and/or the remanng lfetme of the clent may be dfferent from the one predcted on the server sde, the clent wll have to solve another optmzaton problem. Ths tme the clent knows the actual channel state and the adopted parameters of encodng and transmttng on the server sde and has more accurate and up-to-date nformaton about ts own remanng lfetme; therefore, the clent can optmze ts energy consumpton more effectvely and thus obtan and enforce the recepton and decodng parameters whch are agan looked up from the polcy parameter table. The aforementoned polcy optmzaton problem, whch nvolves a herarchcal varable determnaton process, s a form of mult-level optmzaton problems known as tackelberg game [] as detaled next. 4 A Game-theoretc Formulaton for the Dynamcally-varyng hannel ondton 4. Background In hs monograph about market economy [7], H. V. tackelberg used a herarchcal model to descrbe real market condtons. Hs model captured the scenaro n whch dfferent decson makers attempt to make the best decsons n a market wth respect to ther own, generally dfferent, utlty functons. Generally speakng, these decson makers cannot determne ther course of acton ndependently of each Obvously, more elaborate channel estmaton technques may be employed to mprove the selecton process, but ths smple channel predcton scheme serves our purpose of llustratng the general approach. opt other; rather, they are forced to act accordng to a certan herarchy. onsder a smple case of such a problem where there are only two actve decson makers. The herarchy classfes these two decson makers nto a leader, who acts ndependently of the market, and a follower, who has to act n a dependent manner. The leader s able to dctate the sellng prces or to overstock the market wth hs products, but n makng hs decsons, he has to antcpate the possble reactons of the follower snce hs proft strongly depends not only on hs own actons but also on the response of the follower. On the other hand, the choce of the leader nfluences the set of possble decsons as well as the objectves of the follower who n turn must react to the selectons of the leader. The aforementoned problem can mathematcally be formulated as follows: Let X and Y denote the set of admssble strateges x and y of the follower and of the leader, respectvely. Assume that the values of the choces are measured by the means of the functons fl ( xy, ) and ff ( xy, ), denotng the utlty functons of the leader and follower, respectvely. Then, wth the knowledge of the selecton y of the leader, the follower can select hs best strategy x ( y ) so that hs utlty functon s mnmzed on X: { F } x( y) Ψ L( y) = Argmn f ( x, y) x X (5) x Beng aware of ths selecton, the leader solves the tackelberg game [7] for computng hs best selecton: y { L (, ), L ( ) } Argmn f x y y Y x Ψ y (6) It s worth notng that the solutons to the tackelberg game are dfferent from the Nash equlbrum ponts, due to the specal herarchy that s mposed on the players. In Nash equlbrum soluton all players have the same level of herarchy and make decsons smultaneously, but n a tackelberg game the decsons are made one after the other, followng certan rules. In general, n an n-player tackelberg game all players n same herarchy level acheve the Nash s equlbrum pont, but ths s not true for players from dfferent levels of herarchy. 4. Applcaton to treamng Vdeo In our context, the follower and the leader become the clent and the server, respectvely. trategy x for the clent s the adopton of a specfc vector of truncaton lengths (TL s) for the sub-carrers and an operatng frequency for the decoder mage processng core, f, and therefore, { (,,,, ) :, n } X = TL TL L TL f TL TL f F, where n s the number of sub-carrers n the Orthogonal Frequency Dvson Multplexng (OFDM) sgnal, TL denotes the set of all (feasble) TL s for the adaptve Vterb decoder, and F s the set of feasble frequences for the mage processng core. trategy y for the transmtter s a choce of specfc overall

6 transmsson power level and a set of modulaton levels for the dfferent sub-carrers and operatng frequency for the encoder mage processng core, and therefore, ( /,,,,, Y = P R b b L b f ) : b ML, P PL, f F. { Tx s n Tx } where ML and PL denotes the sets of (feasble) modulaton levels for each sub-carrer and avalable power levels for sgnal transmsson. These sets are known from chpset specfcaton or the standard protocol supported by the chpset. Note that ths formulaton can easly be extended to take nto account dfferent transmt power levels for each sub carrer. Ths case s not explored here because t would requre multple output amplfers (one per sub-carrer) n order to support ndependently controlled dfferent power level per sub-carrer. Ths s qute expensve from mplementaton pont of vew. mlar to the case of fxed and known channel condtons, the overall objectve of the clent-server game s to ensure that they acheve an acceptable level of performance whle maxmzng the overall vdeo servce tme. Notce that the vdeo servce s termnated as soon as any one of the server or the clent exhausts ts energy source. The way ths optmzaton problem s solved s that the server and the clent take turn at the begnnng and end of each tme slot. The server s goal s to mnmze the overall energy consumpton of the clent-server system whereas the clent s objectve s to make sure that t wll not exhaust ts energy source any sooner than the server does. In ths way, ths two-player game results n extendng the overall system lfetme by frst mnmzng the energy consumpton and then by ensurng that no one des earler than the other. Detals are explaned below. In tme slot t, the clent (follower) uses the absolute value of the dfference between ts expected lfetmes and that of the server as the cost functon, ff ( xy, ). Obvously, the clent must meet maxmum energy consumpton and constrants under the receved NR value, whch s n turn dctated by the choce of the server parameters, Y ˆ. The clent knows the expected lfetme of the server, L, as a result of the last data transmsson. It must adopt clent parameters (.e., TL and f ) so as to mnmze ff ( xy, ). Notce that L was calculated by the server at tmeslot t- as the rato of the server s remanng energy level to the server s aggregate energy depleton rate. nce the OFDM symbols are transmtted at a constant rate, we can factor the power coeffcents and drop the constant values of equatons () and (). The optmzaton problem n the clent may then be stated as follows: arg mn Xˆ Et, ˆ : ˆ ˆ ˆ, ˆ n X AX + BY REQ X TL F L t (7) where E and t Lt denote the remanng energy value of the clent at tme t, and the expected lfetme of the server, whch was receved at tme t-. The rato of E to t Lt sgnfes the power dsspaton target for the clent n order for t to survve untl the end of the server s expected lfetme. X ˆ refers to the clent s strategy. s the coeffcent vector for computng the power consumpton (.e., per-frame energy consumpton) of the clent,, from the TL s and the f value,.e., Yˆ P t P t =, Xˆ. Where ab, s used to represent the nner product of vectors a and b. The clent s objectve n ths optmzaton step s to fnd ˆX such that ts actual power consumpton becomes as close as possble to ts target power dsspaton. The clent must however, do ths optmzaton under approprate maxmum energy consumpton and constrants. These constrants are wrtten n matrx-vector form as: ˆ ˆ ˆ AX + BY REQ where Y ˆ refers to the server strategy and A and B denote the coeffcent matrces that account for the channel condtons and per-frame energy consumptons of the basc buldng blocks of the clent. s a vector consstng of an upper bound on overall energy consumpton, Emax, and the requred values as shown below: ˆ REQ ( n+ ) ( n+ ) k P k PMU k P. MU. MU k P.. M R L U R R R P f k P α. MU L R A= α L = M M M k PMU M M O. R α L n P f ( n+ ) ( n+ ) L P f Emax β L β L REQ ˆ = B = M M M O M M M L β BE R n n α and β (8) are emprcal coeffcents for lnear estmaton of for the th sub-carrer n terms of the modulaton level and the truncaton length of the decoder. P f s the lnear coeffcent, whch models frequency vs. power characterstcs of the mage processng core. The server (leader), on the other hand, attempts to mnmze the overall energy consumpton of the clent-server system gven the channel condtons provded by the clent. The optmzaton problem s mathematcally formulated as: { ˆ ˆ ˆ ˆ ˆ n + ˆ Ψ ˆ L } arg mn, X, Y : DY REQ, Y PL ML F, X ( Y) where s the vector shown below., Y ˆ (9) sgnfes the power consumpton (.e., per-frame energy consumpton) of the server. D s the coeffcent matrx for lnear estmaton of

7 the VQ and n terms of the NR and the modulaton level. REQ ˆ s a vector representng the mnmum requrements for the VQ and. The ˆ ˆ DY REQ constrant ensures that a mnmum vdeo qualty and for each frame s acheved. ( n+ ) ( n+ ) L mn VQ ω δ L REQ ˆ = = L D = ω δ M M O M M M M n ωn L B P δ f ω and δ ERn κ f () M are the estmated channel transfer functon and the estmaton coeffcent for sub-carrer. Notce that when solvng equaton (9), the optmzaton varables are ˆX and Yˆ, that s, the server estmates the best strategy that the clent may put n practce n response to the server s strategy and then based on ths estmate of the clent s polcy, the server goes on to determne both the optmum values of the server and the clent parameters. It wll then mplement the server polcy, but wll not report or n any way attempt to enforce the clent polcy. The clent n ts own turn wll determne ts optmum polcy parameters as explaned prevously. 5 Expermental Results We mplemented an MPEG-4 FG streamng system on a hgh performance testbed. The hardware used s the Intel s Xscale processor, whch supports 9 dfferent frequences from MHz to 733MHz. A D/A converter was used as a varable operatng voltage generator to control the reference nput voltage to a D-D converter that supples operatng voltage to the PU. Inputs to the D/A converter were generated usng customzed PLD logc. When the PU clock speed s changed, a mnmum operatng voltage level should be appled at each frequency to avod a system crash due to ncreased gate delays. In our mplementaton, these mnmum voltages are measured and stored n a table so that these values are automatcally sent to the varable voltage generator when the clock speed changes. Voltage levels mapped to each frequency are dstrbuted For the software work, Mcrosoft reference MPEG-4 FG encoder/decoder was modfed to ft our purpose. Two generated bt-streams of QIF vdeo sequence wth 5 frames, a base layer and a FG enhancement layer wth 5 bt-planes (bp~bp4), are splt nto packets wth sze of 56-byte. RTP/RTP on UDP was used as a network protocol between the server and the clent. Both the server and the clent were equpped wth an IEEE 8.b wreless LAN (WLAN) card. Energy consumpton of the WLAN nterface was measured by usng a data acquston (DAQ) system. In order to smulate the system, the mulnk 5. envronment from Matlab 6.5 Release 3, was used. To model a mult-path fadng channel, a parallel combnaton of Raylegh and Rcan fadng propagaton channels was used [9]. The maxmum Doppler shft and the spreadng factor of the Rcan fadng channel were set to 4Hz and, respectvely. In order to take nto account the effect of mult-path fadng, three dfferent paths wth delays of us, 3us, and 5us and gans of 3,, and were consdered n Raylegh propagaton channel. The characterstcs of ths channel were smulated and recorded for duraton of 48 frames,.e., fve frames/second. To produce the channel probablty dstrbuton, h(ω ), ths data s used. Then these probablty values are fed nto the tackelberg game for polcy desgn. To show the effectveness of our approach three dfferent scenaros were smulated and compared wth each other. In scenaro number one, for each tme slot, after the detecton of channel condtons we assgned enough energy to the server and the clent to support the specfed average vdeo qualty. In scenaro number two, average channel condton was used to determne the requred energy n each tmeslot, and fnally n scenaro three our adaptve algorthm was used to calculate the requred energy for each tmeslot. Fgure 3 shows the lfetme versus requred average vdeo qualty graphs for scenaros and 3. Total ntal energy for ths experment s set to 6 J. Accordng to ths graph, our approach ncreases the system lfetme by as much as % for hgh values of the requred vdeo qualty. On average, the proposed method ncreases the system lfetme by more than 5% over the whole range of vdeo qualtes. Fgure 4 shows the comparson between scenaros and 3, snce n scenaro, the assgned energy to each frame s fxed, and s selected accordng to the average channel behavor, the system consumes extra energy to produce hgher vdeo qualty, and hence, has a lower system lfetme. #frames) fe tme ( L cenaro vs. cenaro 3 cenaro3 cenaro Vdeo Qualty (%) Fgure 3. Lfetme comparson between cenaro and 3

8 Fgure 4(a) shows comparson between the system lfetmes for scenaros and 3. It s clear that system lfetme s sgnfcantly ncreased for scenaro 3 where we employed the proposed dynamc polcy approach. Notce that the average vdeo qualty was mantaned above the requred value. However, for scenaro, the average vdeo qualty s unnecessarly mproved, whch may be OK f there was no energy dsspaton overhead (cf. Fgure 4(b)). 6 oncluson Moble vdeo streamng system energy consumpton s modeled and analyzed. Usng ths model an adaptve approach to energy assgnment to each frame s proposed. The proposed approach guarantees the mnmum vdeo qualty for all frames, and also meets a requred average vdeo qualty over lfetme of system. Actual expermental data s used to extract the proposed models parameters and based on these parameters smulaton are setup to show the effectveness of the proposed approach. Based on ths approach the overall system lfetme s ncreased by %. ystem lfetme (#frames) Actual Average Vdeo Qualty (%) Requred Average Vdeo Qualty (%) (a) ystem lfetme cenaro 3 cenaro cenaro 3 cenaro Requred Average Vdeo Qualty (%) (b) Vdeo qualty Fgure 4. omparsons between cenaros and 3 References [] W. L, Overvew of Fne Granularty calablty n MPEG-4 Vdeo tandard, IEEE Trans. On rcuts and ystems for Vdeo Technology, Vol., No. 3, Mar.. [] M. van der chaar, H. Radha, and. Dufour, calable MPEG-4 Vdeo odng wth Graceful Packet-loss Reslence over Bandwdth-varyng Networks, Proc. of the IME, vol.3, pp ,. [3] R. ohen and H. Radha, treamng Fne-Graned calable Vdeo over Packet-Based Networks, Proc. of the GLOBEOM. IEEE, pp.88-9,. [4] R. Yan, F. Wu,. L, and R. Tao, Error reslence methods for FG vdeo enhancement btstream, The Frst IEEE Pacfc-Rm onference on Multmeda, Dec. 3-5, ydney, Australa. [5] J. Proaks, Dgtal ommuncatons, McGraw-Hll, 3rd Edton, 995. [6] [7] K. ho, K. Km, and M. Pedram, Energy-aware MPEG-4 FG treamng, Proc. of the Desgn Automaton onference, pp.9-95, 3. [8] T. Perng, T. Burd, and R. Broderson, The smulaton and evaluaton of dynamc voltage scalng algorthms, Proc. of Int l ymp. on Low Power Electroncs and Desgn, pp.76-8, 998. [9] Wreless propagaton bblography, wctg/manet/wrelesspropagaton_bblog.html [] T.M. over, J.A. Thomas, Informaton Theory, nd Edton, John Wley & ons Inc., New York, 99. []. Appadwedula, M. Goel, N.R. hanbhag, D.L. Jones, and K. Ramchandra, Total ystem Energy Mnmzaton for Image Transmsson, Journal of VLI gnal Processng ystems Feb.. [] tephan Dempe, Foundatons of Blevel Programmng, Kluwer Academc Publshers, Boston,. [3] R. Hennng and. hakrabart, Low-power approach to decodng convolutonal codes wth adaptve vterb algorthm Approxmatons, n Proc. of Proc. of Int l ymp. on Low Power Electroncs and Desgn, pp. 68 7, Aug.. [4] F. han and D. Haccoun, Adaptve Vterb Decodng of onvolutonal odes over Memory less hannels, IEEE Trans. on omm., Vol. 45, No., pp , Nov [5]. wamnathan, R. Tesser, D Geockel, and W. Burleson, A dynamcally Reconfgurable Adaptve Vterb Decoder, n Proc. of the FPGA onf., Monterey, alforna, Feb.. [6]. F. Ln and J. B. Anderson, M-Algorthm Decodng of channel convolutonal odes, n Proc. of Prnceton onference of Informaton c. and ystem, pp , Prnceton, NJ, Mar [7] H.v. tackelberg, Marktfrom und Glechgewcht, prnger-verlag, Berln 934. engl. Transl. The theory of the Market Economy, Oxford Unversty Press, 95.