Dynamic Power Management in a Mobile Multimedia System with Guaranteed Quality-of-Service

Transcription

1 Dynmc Power Mngement n Moble Multmed System wth Gurnteed Qulty-of-Servce Qnru Qu, Qng Wu, nd Mssoud Pedrm Dept. of Electrcl Engneerng-Systems Unversty of Southern Clforn Los Angeles CA Outlne! Introducton Overvew of dynmc power mngement Defnton of Qulty of Servce (QoS)! System modelng GSPN bckground Non-exponentl dstrbuton System modelng! Optmzton technque Buffer estmton Polcy optmzton! Expermentl results! Conclusons

2 Introducton! System-level dynmc power mngement Power control of ll system components nd resources Dynmc chnge of the system power stte whle meetng globl performnce constrnt! Lmttons of the prevous works The only performnce constrnt tht hs been consdered s the verge wtng tme of request Servce requests hve been collected from trdtonl pplctons such s fle ccess, key bord ccess, etc. Qulty of Servce (QoS)! QoS: The set of quntttve nd qulttve chrcterstcs of dstrbuted multmed system tht cpture the noton of "user stsfcton" wth the multmed dt presented to hm/her Network

3 QoS Prmeters! QoS s often expressed by set of trget prmeters! The three most mportnt QoS prmeters re: Dely (D): Tme ntervl between the moment dt unt s receved (nput) nd the moment t s sent (output) Jtter (J): Vrton n dely vlues for dt unts n gven nput strem Loss rte (L): Frcton of dt unts tht s lost durng the dt trnsport Globl QoS Mngement! User requests n end-to-end QoS level! Globl QoS mnger lloctes QoS for ech system component! Locl QoS mnger controls the llocton nd stte of the locl resources! The clent system needs power nd QoS mngement (PQM)

4 PQ Mnger! The PQ mnger performs both power nd QoS mngement Determne the PQ mngement polcy tht results n the mnmum power dsspton whle meetng user specfed QoS constrnts (D, J, L) Dtbse Locl QM Server Trnsport System Bndwdth, etc Locl QM Resources Locl QM Clent Globl QM PQM Polcy Optmzton Workflow System Modelng n Controllble GSPN Trnsformton from Controllble GSPN to Controllble CTMDP Buffer Sze Estmton LP-Bsed Polcy Optmzton

5 Bckground: GSPN Prmtves! Plce: condton or stuton! Token Mrkng m(p): #of tokens n p System stte m! Trnston: event Tmed nd mmedte! Input rc: I (t, p) t p, p t! Output rc: O (t, p) t p, p t! Condton Gte: G! Cse: uncertnty m(p on )=0, m(p stby )=1, m(p off )=0, m(p queue )=0 m = [0, 1, 0, 0] P on t process t swtch_on t swtch_off P queue G: m(p queue )=0 P stby P off GSPN Enblng nd Frng Rules! Trnston t s enbled n mrkng m exctly f p t, m(p) I (t, p) nd condton for ny gte G tht s on n nput rc s true! Frng of t Removes I (t, p) tokens from t Deposts O (t, p) tokens nto t t swtch_on P on G: m(p queue )=0 t process P queue P stby t swtch_off α =0.5 β =0.5 P off p on p stby p off p queue (pr = 0.5) (pr = 0.5)

6 GSPN Enblng nd Frng Rules (cont.)! A tmer s ssocted wth tmed trnston t When t s enbled, tmer s set to rndom vlue ccordng to the probblty dstrbuton functon ssocted wth t nd strts countng down When the tmer reches 0, t fres nd resets the tmer! An mmedte trnston lwys hs hgher prorty thn tmed trnston! Mrkng types: Tngble mrkng: no mmedte trnston s enbled Vnshng mrkng: t lest one mmedte trnston s enbled F(t) Controllble GSPN! A controllble GSPN s GSPN where the cse probblty of free-choce conflct mmedte trnstons cn be controlled by outsde commnds Cn be trnsformed to controllble CTMDP Need to fnd the set of commnds (nd hence, cse probbltes) tht mnmze some cost functon t swtch_on t process P stby P on P queue Controllble probblty t swtch_off P off

7 PQ-mnged MM Clent System MM Strem MM Buffer Scheduler Locl pplcton QoS constrnts Request Queue Schedulng Control Servce Provder Power Mode Control Locl PQ Mnger! The nput nter-rrvl tmes of multmed (MM) request generlly follow non-exponentl dstrbuton! The prorty of the MM request s hgher thn tht of the locl (norml) pplcton request! J nd L constrnts re only ppled to the MM pplctons Top Level GSPN Model MMSR&SQ MM Strem MM Buffer SP Scheduler Request Queue Locl pplcton Locl SR & SQ QoS constrnts Schedulng Control Servce Provder Locl PQ Mnger Power Mode Control

8 GSPN Model for the MM SR & SQ Stge 3 Stge 2 Stge 1 P MM1 µ 2 β 2 P MM2 µ 3 µ 1 β 1 α 1 α 2 P MMBuf G MM :m(p MM1 )+m(p MM2 )=0&m(P MMBuf ) < MM buffer sze! Use stge method to pproxmte the nonexponentl nter-rrvl tme of MM request; n ths exmple, r = 3 GSPN Model for the Locl SR & SQ T norm P SQ G norm : m(p SQ ) < SQ cpcty! Assume tht the nter-rrvl tme of locl requests follows n exponentl dstrbuton

9 GSPN Model for the SP T 2s T s2 P 2s P dle (,MM) T decson () P decson () T decson (s) P decson (s) P s2 PMMBuf T strt P (,MM) work T process (,MM) T redecson P dle (s) T vnsh P chngng P SQ P dle (,norm) P work (,norm) T process (,norm) G: m(p dle (s))=0 P sttus (p_mode, _mode)! Sttus: dle, busy, trnston, decson! p_mode: ctve, sleepng, etc.! _mode: MM pplcton, locl pplcton Cost Defnton! Rte cost: d r, j r, l r, ld r, pow r P MMBuf d r = #tokens j r = ( #tokens - Ave(#tokens) ) 2 l r = 1 when #tokens = MM Buffer sze P SQ ld r = #tokens P sttus (power_mode, pplcton_mode) pow r = power consumpton of SP n ts current stte! Impulse cost: ene ene j : Energy needed for the SP to swtch from stte to stte j

10 PQM Polcy Optmzton Workflow System Modelng n Controllble GSPN Trnsformton from Controllble GSPN to Controllble CTMDP Buffer Sze Estmton LP-Bsed Polcy Optmzton Trnsformton from GSPN to CTMP! A stte n CTMDP corresponds to tngble mrkng n GSPN! A trnston n CTMDP corresponds to tmed ctvty n GSPN! The rte cost of stte of CTMDP s the sum of ll the rte costs of the plces n the correspondng mrkng n GSPN! The trnston cost s CTMDP the mpulse cost of the correspondng ctvty n GSPN Serch the rechblty set of GSPN Fnd trnston rte between ech stte nd form the genertor mtrx of CTMDP Clculte the cost nd mpulse cost

11 PQM Polcy Optmzton Workflow System Modelng n Controllble GSPN Trnsformton from Controllble GSPN to Controllble CTMP Buffer Sze Estmton LP-Bsed Polcy Optmzton Buffer Sze Estmton! Too lrge buffer sze s unnecessry! Too smll buffer sze wll overconstrn the system Buffer Sze 4 6 (D,J,L)=(1.5,0.9,0.02) D J L Power ! Gven some buffer sze, the performnce metrcs D, J nd L re dependent on ech other Gven ny three, we cn estmte the fourth one We re nterested n the mnmum buffer sze tht s needed to vod overconstrnng the system

12 Buffer Sze Estmton Mn. n n = D =1 n 2 ( D) p = =0 n p = 1 = 0 p n L 0 p 1, =1,, n Subject to: p A bound on the mnmum requred buffer sze: N=Mx(n 1,n 2,n 3 ) - No overconstrnts f n N J (n -2) 2 L J+D 2 4 D+3 (n +1) (n 2) L D 2 (n 2) (n 1) L D 2 3 D+2 +J n rel vlue estmted vlue D PQM Polcy Optmzton Workflow System Modelng n Controllble GSPN Trnsformton from Controllble GSPN to Controllble CTMP Buffer Sze Estmton LP-Bsed Polcy Optmzton

13 PO Problem Formulton Mn } τ { x Subject to: x ( pow + ene p ) j j x j p j = x 0 j j = x τ 1 < x q _ MMBuf τ x 2 ( q _ MMBuf D) τ < J x q _ SQ τ < x l 0 j D j D j p j : probblty tht the next system stte s j τ : expected durton of tme tht the system wll be n stte x : frequency tht the stte of the system wll be nd cton wll be tken; Note tht: x τ p pow : power consumpton n stte q_mmbuf : number of unprocessed dt n the MM buffer ene j : the energy needed for system to swtch from stte to stte j Lner Approxmton of Jtter! The exct jtter: MMBuf x q _ x q _ MMBuf Non-lner expresson of x! Lner pproxmton of jtter: 2 x ( q _ MMBuf D) τ ( 2 τ ) τ () Lner expresson of x! Theorem: For ny set of { x }, f (b) s smller thn J then () s smller thn J. For ech polcy, f the pproxmted jtter stsfes the gven constrnt then the rel jtter lso stsfes the gven constrnt (b)

14 Expermentl Results! System setup SP hs two power modes: hgh power nd low power In the PD_optmzed system, we end up overconstrnng dely n order to stsfy the jtter constrnts Power Consumpton (mw) MM locl dle Hgh Power Low Power Hgh Power Low Power Servce Speed (ms) MM locl dle (1,1,0.1%) 18 (1,1.5,0.1%) (3,1,0.1%) 16 (3,1.5,0.1%) 14 (5,1,0.1%) (5,1.5,0.1%) Percentge Power Svng of PQM over PDM Conclusons! We ntroduced complete modelng technque bsed on controllble GSPN wth cost tht cptures the behvor of bttery-powered multmed clent system! We showed how to obtn the PQ-optml polcy bsed on ths stochstc mthemtcl frmework! Ths s the frst power mngement polcy tht consders jtter nd loss rte s well s the dely! Expermentl results demonstrted tht the PQoptmzed plces re more power-effcent thn the PD-optmzed polces under the sme D, J nd L constrnts