Distributed Uplink Scheduling in EV-DO Rev. A Networks

Transcription

1 Dstrbuted Uplnk Schedulng n EV-DO ev. A Networks Ashwn Srdharan (Sprnt Nextel) amesh Subbaraman, och Guérn (ESE, Unversty of Pennsylvana)

2 Overvew of Problem Most modern wreless systems Delver hgh performance through tght control of transmssons by the Base Staton (whch devces, when & at what power) Most modern wreless devces un a broad range of applcatons wth dfferent communcaton needs (voce, vdeo, web, emal, SMS) Centralzng all decsons at the base staton lacks flexblty and scalablty Latest wreless standards nclude mechansms for partally delegatng transmsson decsons to devces But there s a cost n gvng devces autonomy n makng ndependent transmsson decsons? Sub-optmal resources sharng can mpact overall throughput How bg s the problem? What polces/mechansms to best mtgate those effects? 5/3/007 Networkng Atlanta

3 System Overvew Base Staton Internet Gateway Internet Base Staton Controller Moble Swtchng Center Telephone Network 5/3/007 Networkng Atlanta 3

4 Our Focus 5/3/007 Networkng Atlanta 4

5 Overvew of esults Assessng the mpact of ndependent (uplnk) user transmssons Saturated, homogenous users andomzed polces (transmsson probablty p) Optmal value for p wth sgnfcant mpact on throughput Threshold behavor as a functon of system load ealzng optmzed dstrbuted transmssons n token bucket controlled systems Selectng transmsson probabltes to approxmate optmal polces under bucket constrants 5/3/007 Networkng Atlanta 5

6 Outlne of Talk A short prmer on wreless transmssons CDMA uplnk EV-DO ev. A operaton Prevous works Modelng dstrbuted transmsson decsons Analyss of randomzed polces Emulatng optmal polces Token-bucket controlled systems Extensons of results and future work 5/3/007 Networkng Atlanta 6

7 Overvew of CDMA Uplnk CDMA uplnk s nterference lmted Each user has a spreadng orthogonal code Allows smultaneous transmssons However, users nterfere due to mult-path effects Users can select among multple (dscrete) transmsson rates Control loop based on plot sgnal equalzes channel among users Transmtted power s proportonal to plot strength AND selected rate 5/3/007 Networkng Atlanta 7

8 Uplnk Operaton Plot P transmtted by devce =1,...,n+1 Plot sgnals are power controlled by BS to all be receved wth the same target SIN 1/Ф 1 GlossP σ = GlossP = Δ =, = 1, K, n j φ σ + θ G P φ nθ + Plot loss j Plot j G loss : Path loss; θ Plot : Orthogonalty factor; σ : Nose User transmt power = P TxTP[] : Target data rate from dscrete set TxTP[] : Proportonalty factor relatve to Plot User spends TxTP[] power tokens to transmt at rate 1 5/3/007 Networkng Atlanta 8

9 Sample TxTP[] Values Target Data ate TxTP[] db kbps kbps kbps kbps kbps /3/007 Networkng Atlanta 9

10 5/3/007 Networkng Atlanta 10 CDMA Uplnk Interference Model Plot j j D j j j D j loss D loss n P TxT P TxT G SIN P TxT P P W G P G P G G SIN θ φ σ θ σ θ θ σ Δ = Δ + Δ = = = + =, ] [ ] [ ) ( ) ( ] [ ) ( : ) ( :, ) ( ) ( ) ( ) ( and Processng Gan Data orthogonalty factor No Channel Effects (Perfect Power Control) Interferences from other users The hgher the rate a user chooses, the more nterference t creates!

11 Our Problem SIN G( ) TxT P[ ] Δ ( ) =, = 1, K, n + 1 σ + θ TxT P[ ] Δ j Users make ndependent transmsson and rate selecton decsons Greedy behavor by ndvdual users can affect overall performance What gudelnes to mtgate negatve mpact of ndependent decsons j 5/3/007 Networkng Atlanta 11

12 Prevous Work Extensve work on rate allocaton and power control Assumes contnuous transmsson (no schedulng). Schedulng n CDMA ad-hoc networks Assumes synchronzaton, contenton resoluton. Closest work that of [3], [4] Schedulng n cellular CDMA. Solves centralzed global allocaton numercally. 5/3/007 Networkng Atlanta 1

13 Our Intal Model Homogenous, unconstraned users All users (n+1 users n a sector) employ the same polcy Users always have data and are able to transmt whenever the polcy schedules a transmsson Probablstc On-Off transmsson polcy Transmt at rate n a slot wth probablty p Transmt power s therefore 0 wth probablty 1-p and ~TxTP[] wth probablty p Smple but useful model Smlar to Aloha But wth a contenton model based on soft nterferences (CDMA) rather than collsons Questons At what rate should a user transmt? How often (what p value) should a user transmt? 5/3/007 Networkng Atlanta 13

14 Man esults There exsts an optmal p * (maxmzes C ˆ( p ) ) If δ 1 then p * =1 φ nθ δ = Plot If δ < 1 then p * < 1 θ TxT P[ ] In both cases p* satsfes the followng equalty n 1 n j n j 1 p* (1 p*) = + δ + 1 j= 0 j j ( n 1) p* +δ Wth few (many) users, and/or low (hgh) target rate, users should transmt (n)frequently Hgher target rates always acheve hgher * * throughput,.e., C ˆ( p, ) > Cˆ( p, f > 1 1 ), In the absence of other constrants 5/3/007 Networkng Atlanta 14 1

15 Impact of δ 30 Throughput (LINEA MODEL) Users 45 Users Transmsson Probablty 5/3/007 Networkng Atlanta 15

16 Hybrd Slotted/CDMA Throughput (LINEA MODEL) Bounded model: C( ) SIN( ) = mn, S0 Lnear ate Bounded ate Slot Dvson Throughput (BOUNDED ATE) Transmsson Probablty 5/3/007 Networkng Atlanta 16

17 Dstrbuted Control Token bucket mechansm avalable n EV- DO ev. A and HSUPA to control devce transmssons Token bucket depth σ and token fll rate ρ are controlled by Base Staton A devce needs TxTP[] tokens to transmt at rate Amed at lmtng peak and average power to satsfy farness and QoS constrants Queston: How does the presence of a token bucket affect the choce of good transmsson decsons by devces? 5/3/007 Networkng Atlanta 17

18 Accountng for Token Buckets Gven a token bucket confguraton (σ,ρ) What are the optmal p* and K values? Two-step formulaton 1. Account for mpact of token bucket on transmsson decsons Transmssons condtoned on havng at least K tokens. Explore possble combnatons of p and K values Note that optmalty of hgher rates need not hold any more because of token constrants (token effcency) 5/3/007 Networkng Atlanta 18

19 Token Effcency 3.8 Wth 4 users transmsson at 153.6kbps yelds a hgher throughput but a lower token effcency than transmsson at 76.8kbps Token Effcency (BOUNDED ATE) kbps kbps Transmsson Probablty 5/3/007 Networkng Atlanta 19

20 Impact of Token Bucket Token Bucket parameters: σ = 1.5dB; ρ = 7dB Throughput(BOUNDED ATE) More frequent transmssons at 76.8kbps yeld a better throughput because of hgher token effcency 76.8 kbps kbps Transmsson Probablty Condtonal Transmsson Probablty 5/3/007 Networkng Atlanta 0

21 Analyss vs. ealty Token Bucket: σ = 1.5dB; ρ = 7dB ate (kbps) Analyss Smulatons (bounded rate model) p * A C * A p* sm C * sm C sm (p * A ) Expected naccuraces because of bounded rate But actual mpact on throughput s small 5/3/007 Networkng Atlanta 1

22 Extensons & Future Work ecent results Establshed that smlar results also hold for the bounded rate model Characterzed optmum centralzed schedule A benchmark aganst to compare dstrbuted polces A combnatoral problem because of dscrete rate values Extensons Investgatng the mpact/use of token bucket for ts orgnal purpose, namely, servce dfferentaton ate vs. delay performance targets 5/3/007 Networkng Atlanta

23 elevant eferences 1. P. Venktasubramanam, S. Adreddy, and L. Tong, Opportunstc ALOHA and cross-layer desgn n sensor networks. Proc. IEEE MILCOM, Boston, MA, October P. Venktasubramanam, Q. Zhao, and L. Tong, Sensor networks wth multple moble access ponts. Proc. 38th Annual Conference on Informaton Systems and Scences, Prnceton, NJ, March K. Kumaran, L. Qan, Uplnk Schedulng n CDMA Packet-Data Systems, INFOCOM Cruz, A. Santhanam, Optmal outng, Lnk Schedulng and Power Control n Mult-Hop Wreless Networks, INFOCOM /3/007 Networkng Atlanta 3