Combining age-based and channelaware scheduling in wireless systems

Transcription

1 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Combiig age-based ad chaelawae schedulig i wieless systems Samuli Aalto ad Pasi Lassila Netwokig Laboatoy Helsiki Uivesity of Techology {Samuli.Aalto Pasi.Lassila}@tkk.fi (0)

2 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia HSDPA/HDR systems Dowlik tasmissios BS tasmits to exactly oe use i a time slot with full owe Schedulig: BS decides allocatio of time slots fo diffeet uses taffic Fo examle: oud obi schedulig (0)

3 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Flow-level esective We coside elastic taffic Coesods to uses efomig web sufig cosistig of file tasfes Elastic meas that alicatios toleate vaiatios i istataeous ates I the dyamic settig file tasfes (o flows) aive adomly (Poisso aivals) ad have adom sizes (tyically heavy tailed) Pefomace exessed as mea file tasfe delay o thoughut Uses oly cae about the total time to tasmit/eceive the comlete file Coectio back to time-slot level To tasmit a tyical file equies may time slots Diffeet taffic model fom acket level aoaches with fo examle i.i.d. aivals e time slot c.f. cµ-ule (Stolya et al.) 3(0)

4 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Ootuistic schedulig / size-based schedulig Fast fadig: the ate (o SNR) chages adomly i each time slot (mobility) Ootuistic chael awae schedulig Idea is to exloit the chael vaiatios betwee uses ad give the time slot to uses i a good state (with high ate) Caacity iceases due to schedulig gai Stadad age-based schedules i fast fadig eviomet Idea is to get id of small flows as quickly as ossible to miimize flow delay Deedig o what ifomatio is available we have diffeet olicies SRPT FB (LAS) PS Stadad aoach would utilize kowledge of file sizes (bits) ad mea ate of the uses (ot the istataeous ates) 4(0)

5 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia PF comaed with stadad SRPT ad FB Age-based schedules (SPRT FB) efom bette tha simle PS but gai fom ootuistic schedulig ca be (PF) much geate 0 PF low 8 Paametes E[T] 6 4 PS FB SRPT Rayleigh fast fadig chael Rate is liea i SNR u to maximum ate Symmetic uses PF high Ρ 5(0)

6 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Combiig size/ate ifomatio Poblem: how to combie istataeous size ad ate ifomatio? Difficult oblem otimal solutio is ot kow Ou aoach ad esults Possible to deive may heuistics that combie size/ate ifomatio We assume that sizes obey a cotiuous distibutio (DHR tye) but the ossible chael ates fom a discete set Aalytical esults comaig the otimal olicy ad some heuistics i a simle static settig Simulatios ude heavy taffic to exloe tadeoffs 6(0)

7 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Diffeet ways to utilize size/ate ifo Assumtio: thee is a discete set of ossible use ates Also all uses ca achieve the maximum ate! Two classes of olicies Pioity olicies Idex olicies Pioity olicies Absolute ioity o highest ate Aly size ifomatio to beak ties Geedy aoach fo utilizig the chael Idex olicies Sigle idex value that combies ate ad size ifomatio 7(0)

8 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Pioity olicies Give absolute ioity to highest istataeous ate Idea: utilize the chael maximally If multile flows have same highest ate vaious olicies ae ossible deedig o size ifomatio available SRPT-P seve flow with least amout of bits left aims fo maximum efficiecy FB-P seve flow with least amout of bits seved same as SRPT-P but with oly kowledge of attaied sevice (i bits) RR-P : seve flow with smallest thoughut (attaied sevice / time i system) aims fo iceased faiess 8(0)

9 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Idex olicies PF (Relatively best) Select use k with highest R k / γ k R k = istataeous ate of use k ad γ k = thoughut of use k RB (Relatively best) Select use k with highest R k /E[R k ]; blid olicy with esect to size ifo TAOS (Hu et al. Comute Netwoks 004) Otimal oe ste decisio ule fo imovig basic SRPT olicy M = of jobs i the system X k = emaiig umbe of bits fo use k (SRPT-like ifomatio) Uses ae aked i ascedig ode of X k /E[R k ] (basic SRPT) I k = ak of use k select use k* so that k * = ag mi k ( M I + ) [ R ] k FB-TAOS Relace X k with attaied sevice A k i akig (i.e. seved bits thus fa) k R E k 9(0)

10 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Aalytical study of otimal olicy Assumtios take time slot legth = ossible ates ( mi max ) = ob. that ate is mi (symmetic case) jobs with give size (size / mi = itege) o ew jobs aive Simle discete time decisio oblem Pefomace Total time to seve both jobs util comletio (total comletio time) Objective Comae otimal olicy with stadad PF olicy ad those schedules that use SRPT like ifomatio Refeece schedules: PF ad two best schedules SRPT-L ad TAOS Otimal olicy ca be solved via dyamic ogammig 0(0)

11 (0) HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Otimal olicy Otimal olicy ca be solved by dyamic ogammig Easy to comute v(0**) fo all 4 combiatios of chael states ad the just iteate the above Simila aalysis ossible also fo PF SRPT-L ad TAOS ( ) ( ) ( ) [ ] ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) max max mi max max mi mi mi max max mi max max mi mi mi Mi v v v v v v v v v = seve with ate seve with ate

12 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Comaisos with otimal olicy Delay atio RB max = 6 RB max = 0 SRPT-P TAOS RB max = Paametes mi = = 0.5 (symmetic chaels) chael iitial state = {} Coclusio: TAOS ad SRPT-L vey close to otimal while imovemet ove PF ca be 5-0% (0)

13 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Dyamic simulatios Idea is to study the heavy taffic behavio of the olicies ude diffeet settigs fo the use ates I the settig whee is vey small comaed with time scale of aivals ad deatues Poisso aivals Paeto(.0) file sizes We fix λ= ad vay the sevice times to get diffeet loads Symmetic (= all uses ae idetical) vs. asymmetic settigs use classes equal aival ates i both classes λ = λ = 0.5 symmetic/asymmetic achieved via aameteizatio of ates Diffeet ate sceaios i a i.i.d. chael Case: oly ossible ates small diffeece Case : oly ossible ates lage diffeece Case 3: same set of ates as i HSDPA ( ates) 3(0)

14 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Symmetic case ates low vaiability E[N] Paametes: = = = 0.5 = 0.5 Idex olicy Pioity olicy Fai FB SRPT Ρ* Commets All othe olicies bette tha PF SRPT ad FB like olicies seaate icely Coesodig P-olicy bette tha its TAOS vaiat Ρ 4(0)

15 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Symmetic case ates high vaiability 0 Paametes: = = 0 = 0.5 = 0.5 Commets E[N] Idex olicy Pioity olicy Fai FB SRPT Ρ* All olicies bette tha PF SRPT- ad FB-like olicies seaate icely SRPT-P ad TAOS ealy same FB-P ad FB-TAOS ealy same Caacity limit is highe due to highe schedulig gai Ρ 5(0)

16 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Symmetic case HSDPA ates Paametes: uifom distibutio fo ates E[N] Idex olicy Pioity olicy TAOS FB-TAOS PF Commets Both TAOS olicies wose tha PF P-olicies bette tha PF P-olicies soted as exected (SRPT FB RR) Caacity limit is highe due to highe schedulig gai SRPT-P FB-P Fai-P * 6(0)

17 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Asymmetic cases Much moe comlex dyamics The diffeet aoaches ae ot systematically aymoe bette tha othes (as i symmetic case) Degee of asymmety is also oe abitay aamete Some obsevatios Deedig o the load oe method might be bette/wose tha aothe I tems of faiess the elative olicies behave diffeetly fo low loads ad high loads (o-mootoous behavio) 7(0)

18 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Asymmetic case: ates high asymmety E[N] Idex olicy Pioity olicy Ρ Commets Fai RB FB RB ad FB-TAOS stat isig ucotollably P-olicies ae quite good SRPT Ρ* Faiess measue Idex olicy Pioity olicy FB-P Fai-P PF Ρ Commets SRPT-P TAOS RB FB-TAOS Faiess of RB ad FB-TAOS (ad TAOS) have o-mootoous behavio P-olicies ad PF ae quite stable 8(0)

19 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Asymmetic case: HSDPA ates high asymmety 300 HSDPA ates highly asymmetic sceaio (q = 0.5) SRPT-P E[N] RB SRPT-RB PF 50 ρ* Ρ 9(0)

20 HELSINKI UNIVERSITY OF TECHNOLOGY CLOWN semia Coclusios Flow-level age-based schedulig ca give sigificat efomace beefits eve i a time vayig chael settig Aoaches i the ate ifomatio: ioity vs. elative size ifomatio: SRPT vs. FB Aalysis of otimal olicy is difficult Based o cuet aalysis SRPT-L ad TAOS ae quite close to otimal Dyamic studies with diffeet olicies Symmetic case: tadeoffs ae easie to udestad Asymmetic case: may iteestig heomea occu as a fuctio of load whe comaig the olicies dawig coclusios moe difficult 0(0)