scheduling for multiple antenna wireless systems: capacity ...€¦ · scheduling for multiple...
TRANSCRIPT
Scheduling for Multiple AntennaScheduling for Multiple AntennaWireless Systems: Capacity andWireless Systems: Capacity and
User Performance LimitsUser Performance Limits
Prof. Robert W. Heath Jr.Joint work with Manish Airy and Prof. Sanjay Shakkottai
Wireless Networking and Communications Group (WNCG)Dept. of Electrical and Computer Engineering
The University of Texas at Austin
2
OutlineOutline Introduction to the WNCG @ UT Austin
Scheduling in MIMO ChannelsScheduling in MIMO Channels
User-Level Performance LimitsUser-Level Performance Limits
Limited FeedbackLimited Feedback
Conclusions
3
WNCG at The University of Texas AustinWNCG at The University of Texas Austin Wireless Networking and Communications Group
14 faculty 60 graduate students Networking, wireless, devices
7 Industrial affiliates Metrowerks, Motorola, National Instruments, SBC, Texas
Instruments, Time Warner Cable, U.S. Department of Defense Symposium in October
October 20 - 22, 2004 Wed-Fri before the Info Theory Workshop in San Antonio!
www.wncg.org
4
StudentsStudentsCurrent Students Manish Airy: Scheduling for multi-user MIMO systems Robert Daniels: TBD Antonio Forenza: Channel modeling & antenna design for MIMO systems Caleb Lo: Interference management in MIMO ad hoc networks Bishwarup Mondal: Adaptive feedback / data mining application in comm Muhammad Farooq Sabir: Image and video in MIMO systems Roopsha Samanta: Frame theory & channel quantization Ahmad Sheikh: TBD Taiwen Tang: MAC design, MIMO ad hoc networks Chwan-Ming Wang: Handoff in MIMO-OFDM systems
Graduated Students David J. Love: Limited feedback MIMO communication systems
Starting as an Assistant Professor at Purdue Fall 2004
5
OutlineOutline Introduction to the WNCG @ UT Austin
Scheduling in MIMO ChannelsScheduling in MIMO Channels
User-Level Performance LimitsUser-Level Performance Limits
Limited FeedbackLimited Feedback
Conclusions
6
Scheduling in the MIMO BCScheduling in the MIMO BC
7
Prior WorkPrior Work
Core idea ofCore idea of multiuser multiuser diversity diversity Exploit independence of the channels between multiple users toExploit independence of the channels between multiple users to
obtain diversity and thus better ratesobtain diversity and thus better rates Early work on Early work on multiusermultiuser diversity (not MIMO) diversity (not MIMO)
MultiuserMultiuser diversity on the uplink [KnoHum diversity on the uplink [KnoHum’’95]95] MultiuserMultiuser diversity on the downlink [Tse diversity on the downlink [Tse’’97]97]
Characterizing delay-capacity tradeoffsCharacterizing delay-capacity tradeoffs Infinite file size and fixed users [VisTseLarInfinite file size and fixed users [VisTseLar’’02], [VisVenHua02], [VisVenHua’’03]03] Throughput optimal scheduling [Throughput optimal scheduling [ShaSriStoShaSriSto]]
On limiting scheduling modelsOn limiting scheduling models ““FastFast”” channel limit [Bor channel limit [Bor’’03],[PraVee03],[PraVee’’03],[AirShaHea03],[AirShaHea’’03]03] ““SlowSlow”” channel limit [AirShaHea channel limit [AirShaHea’’03]03]
8
MultiuserMultiuser Diversity & MIMO Diversity & MIMO Fixed number of users
Single user (time sharing) [Tel’95] Multiuser broadcast [VisJinGol’02]
Capacity increase with number of users With multiplexing [HocVis’02] With time-sharing [ChuHwaKimKim’03] With linear receivers (mux / sharing) [HeaManPau’01] In cellular systems [VisVenHua’03] With reduced feedback [GesAlo’03]
Others as well Boche, El Gamal, etc.
9
Motivation: BC Multiplexing AlternativesMotivation: BC Multiplexing Alternatives
How do we allocate substreams to users?
Single-user (time-sharing) spatial multiplexing Multiple data streams to single user Spatially-greedy if single best user
Multi-user transmission Every scheduling instant transmit to multiple users “Collectively” greedy multiplexing if to the “best” users
Average delay as choice of QoS metric Scheduling policy enforces the QoS constraint
10
System ModelSystem Model
i is the index of the i-th user s(t) is the transmitted signal Hi(t) is the Mr x Mt channel matrix for user i
Assumed known perfectly at the MS Assumed known perfectly at the BTS
vi(t) is AWGN
11
Single User MultiplexingSingle User Multiplexing Capacity for user i is given by
Where are the eigenvalues of andis obtained from waterfilling
Spatial greedy (MaxRate) scheduler achieves
12
MIMO BC CapacityMIMO BC Capacity
Sum Capacity of MIMO BC, [VisJinGolSum Capacity of MIMO BC, [VisJinGol’’02]02]
Sum capacity scheduler achieves this rateSum capacity scheduler achieves this rate
13
Multiuser Multiuser Sum CapacitySum Capacity
Ergodic sum capacity increases with users
MaxRate (multi-user diversity)Single user transmissionto rate maximizing channel
MaxSumRateMulti-user transmission tosum rate maximizing channels
Example2x2, Uncorrelated H
0 20 40 60 80 1000
1
2
3
4
FCFSMaxRateMaxSumRate
SNR = 0 dB
Number of users
Normalized Ergodic Capacity
Asymptotically in number of users, ergodic capacity for MaxRate and MaxSumRate scheduling scales identically.
14
Number of Number of ActiveActive Users Users
How many active users account for sum capacity?
2 3 4 5 6 7 8 90
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Number of Active Users
Probability Mass Function
Distribution of Number of Active Users
------- 2x2
------- 4x4
SNR = 0 dB, Total user population K = 20 Example Uncorrelated H, 100 trials Roughly linear increase in
total active users withnumber of antennas
15
Rates of Rates of ActiveActive Users Users
What is the allocated rate to each active user? Overall throughput increases Each user gets lower rate when scheduled
0 0.5 1 1.5 2 2.5 3 3.50
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Distribution of Individual Rates
Rate (Normalized to ergodic channel capacity)
Prob(Rate > x)
2x2, User 12x2, User 22x2, User 34x4, User 14x4, User 24x4, User 34x4, User 44x4, User 54x4, User 6
0 dB SNR
Example Uncorrelated H, 100 trials At 0 dB SNR, with MT
antennas, ~85% of mediansum capacity achieved byMT users
16
OutlineOutline Introduction to the WNCG @ UT Austin
Scheduling in MIMO ChannelsScheduling in MIMO Channels
User-Level Performance LimitsUser-Level Performance Limits
Limited FeedbackLimited Feedback
Conclusions
17
Conflicting objectives
Maximize overall throughput (or sum capacity)
Maintain acceptable user experienced delay
Multi-user sum capacity grows with number of users True for both single- and multi-user greedy scheduling Each user competes with many users for access Each user gets lower allocated rate than single-user scheduling
With an average delay constraint, can we still expect overall capacityincrease with number of users?
User Level PerformanceUser Level Performance
18
File Model and AssumptionsFile Model and Assumptions Arrivals at BTS are variable length files
Assume arrivals are geometric Arbitrary file size distribution Implication is that there are a varying number of users
Other assumptions Symbol period is smaller than all events of interest Need this with certain fluid limits
19
System state S(t) = {N(t), H(t), A(t), F(t)} N(t) is the number of active users H(t) = {H1(t), H2(t), …, HN(t)(t)} is the channel process A(t) = {A1, A2, …, AN(t)} are the arrival times F(t) = {F1(t), F2(t), …, FN(t)(t)} are the remaining file sizes
System evolution in discrete time
Discrete-time System ModelDiscrete-time System Model
20
νν Allocated rate, RAllocated rate, Rii, where i=1, 2, , where i=1, 2, ……, N(t), N(t)
νν State evolutionState evolutionνν FFii(t+(t+ΔΔt) = t) = FFii(t) (t) –– ΔΔt Rt Rii(t)(t)νν If If FFii(t+(t+ΔΔt) = 0 then N(t+t) = 0 then N(t+ΔΔt) = Nt) = N(t)(t) –– 1 1
Allocated RateAllocated Rate
Multi-userMulti-user
Single-userSingle-user
21
νν Scheduling as a user selection processScheduling as a user selection process((drop the time index for notational simplicitydrop the time index for notational simplicity))νν FCFSFCFS
νν SRPTSRPT
νν MaxRateMaxRate
νν MaxSumRateMaxSumRate
Scheduling AlgorithmsScheduling Algorithms
22
Over a large enough time interval system is stationary Follows from stationarity of the channel and arrival process
“Fast” Channel evolution (upper bound) Slot duration → 0 Channel evolution independent every time-slot
channel bandwidth → ∞ / coherence time → 0 “Slow” Channel evolution (lower bound) Slot duration → 0 Channel state fixed for the duration of each job channel coherence time → ∞
Limiting ModelsLimiting Models
23
Limiting Analysis - SummaryLimiting Analysis - Summary
SimulationProcessor sharingMaxSumRate
Preemptive Resume(M/G/1)Processor sharingMaxRate
Exact (SRPT delaycalculations)ApproximateSRPT
PK formula (M/G/1)PK formula (M/G/1)FCFS
“Slow” Channel Limit“Fast” Channel Limit
Limiting anls. -> different limiting queueing models Can calculate user data rate & average job delay
24
Some Details (fast limit)Some Details (fast limit)
SRPT and FCFS calculations are straightforward For MaxRate and MaxSumRate use continuous-
time Markov chain Arrivals are Departures are and are state dependent
25
Some Details (continued)Some Details (continued) Average job delay per user
Steady state distribution of users
Expected file length
26
Average ThroughputAverage Throughput Use uniform convergence of DT Markov chain to
a CT Markov chain [PraVee]
27
Numerical ResultsNumerical Results
Assumptions Downlink 0 dB or 10dB SNR, 2 x 2 MIMO Bandwidth 1.25 MHz Perfect channel knowledge at BTS Web downloads with mean 12 Kbytes
Limiting Scenarios “Fast” channel evolution “Slow” channel evolution
28
““FastFast”” Channel (0dB) Channel (0dB)
0 1 2 3 4 5 6 70
0.5
1
1.5
FCFSSRPT-SimulatedSRPTMaxRateMaxSumRateAverage Delay threshold = 1 sec
"Fast" Channel Variation, SNR = 0 dB
Throughput (Mbps)
Average Delay (seconds)
At 1 sec average file delay:φ
SRPT = 1.2 φFCFSφ
SRPT (simulated) = 1.5 φFCFSφ
MaxRate = 2.7 φFCFSφ
MaxSumRate = 4.1 φFCFS
29
““FastFast”” Channel (10dB) Channel (10dB)
30
““SlowSlow”” Channel (0dB) Channel (0dB)"Slow" Channel Variation, SNR = 0 dB
0
0.5
1
1.5
0 0.5 1 1.5 2
Throughput (Mbps)
Ave
rage
Del
ay (
seco
nds)
FCFSSRPTMaxRateMaxSumRate
At 1 sec average file delay:φMaxSumRate = 1.4 φFCFS
Average Delay = 1 sec
31
““SlowSlow”” Channel (10dB) Channel (10dB)
32
An Asymptotic ResultAn Asymptotic Result
For the For the multiuser multiuser collectively greedy schedulingcollectively greedy schedulingpolicy, apolicy, asymptotically,symptotically,
IntuitionIntuition Channel Channel ““hardeninghardening””, [HocMar, [HocMar’’03]03]
33
DiscussionDiscussion “Fast channel” limit (maximal opportunistic gains)
Both greedy schedulers yield lower delay than the SRPT andFCFS scheduler at all points
“Slow channel” limit (minimal opportunistic gains) At low SNR
MaxRate delay slightly better than MaxSumRate scheduler SRPT delay slightly better than MaxRate Better to serve and remove users than to wait for higher rate
At high SNR MaxSumRate similar as fast channel regime
34
OutlineOutline Introduction to the WNCG @ UT Austin
Scheduling in MIMO ChannelsScheduling in MIMO Channels
User-Level Performance LimitsUser-Level Performance Limits
Limited FeedbackLimited Feedback
Conclusions
35
Limited FeedbackLimited Feedback
Problem StatementProblem Statement
Iterative AlgorithmIterative Algorithm
Numerical ResultsNumerical Results
DiscussionDiscussion
36
Problem StatementProblem Statement
How to achieve an arbitrary rate vector within theHow to achieve an arbitrary rate vector within thecapacity region of the MIMO BC?capacity region of the MIMO BC? Without full channel knowledge at the BTS?Without full channel knowledge at the BTS?
Capacity RegionCapacity Region
Sum Capacity PointSum Capacity PointCovariances Covariances for for this rate vector?this rate vector?
37
Solutions, Prior WorkSolutions, Prior Work
Trivial SolutionTrivial Solution Appropriate time-sharing of scheduling policies thatAppropriate time-sharing of scheduling policies that
allocate rates achieved by using transmit allocate rates achieved by using transmit covariancescovariancescorresponding to the boundary of the capacity regioncorresponding to the boundary of the capacity region((requires full channel knowledgerequires full channel knowledge))
Indirect Solution, [VisVenHuaIndirect Solution, [VisVenHua’’03]03] Solve the uplink rate-constrained problemSolve the uplink rate-constrained problem
((requires full channel knowledgerequires full channel knowledge)) Use duality transformations to get the downlinkUse duality transformations to get the downlink
covariances covariances corresponding to the given rate constraintcorresponding to the given rate constraint
38
Proposed SolutionProposed Solution
Main IdeaMain Idea Start with arbitrary downlink transmit Start with arbitrary downlink transmit covariancescovariances Using Using eigenspace eigenspace rotations, iteratively update transmitrotations, iteratively update transmit
covariances covariances until the rate constraint is satisfieduntil the rate constraint is satisfied Algorithm has a limited feedback interpretationAlgorithm has a limited feedback interpretation
(similar to fast power control for CDMA)(similar to fast power control for CDMA)
Other Other eigenspace eigenspace algorithms are available,algorithms are available,[YeBlum[YeBlum’’03], [PopPopRos03], [PopPopRos’’02]02]
39
Basic RotationsBasic Rotations
Consider the complex space, Consider the complex space, CCMMTT
There are There are MMTTCC22 orthogonal 2-dimensional subspaces orthogonal 2-dimensional subspaces Example, in Example, in CC33 there are 3 orthogonal 2-D subspaces there are 3 orthogonal 2-D subspaces For any For any vv11, , vv22 ∈∈ CC33, there exists , there exists QQ such that such that vv11==QQ vv22
QQ may be expressed in terms of may be expressed in terms ofrotations confined to the rotations confined to the MMTTCC22
2-D subspaces of 2-D subspaces of CC33
A A basicbasic rotation rotation QQ, is such that, is such thatd(d(vv, , QQssvv) ) ≤≤ γγ, s = 1, 2, , s = 1, 2, ……, , MMTTCC22
40
Iterative Iterative Multiuser Multiuser AlgorithmAlgorithm
Assume arbitrary user encoding order (FCFS)Assume arbitrary user encoding order (FCFS) Because of Because of ““dirty paperdirty paper”” encoding, user k experiences encoding, user k experiences
interference from users with index j<kinterference from users with index j<k Estimate interference covariance for user k, Estimate interference covariance for user k, σσ22
int,kint,k RRkk
Estimate effective channel for user k, Estimate effective channel for user k, __kk=(=(II + + RRkk))-1/2-1/2HHkk
Waterfilling Waterfilling over over __k k yields optimum transmit eigenvectors andyields optimum transmit eigenvectors andpower allocation per eigenvectorpower allocation per eigenvector
Rate constraint imposed by choosing appropriate Rate constraint imposed by choosing appropriate waterfilledwaterfilled““levellevel””
41
Limited Feedback InterpretationLimited Feedback Interpretation
The transmitter broadcasts a The transmitter broadcasts a ““dirty paperdirty paper””encoded encoded ““interference estimationinterference estimation”” waveform waveform(similar to channel training periods)(similar to channel training periods)
User k computes effective whitened channelUser k computes effective whitened channel Feedback per user is Feedback per user is MMTTCC2 2 + M+ MT T bitsbits
1 bit required to indicate choice of 1 bit required to indicate choice of QQss,, s=1, 2, s=1, 2, ……, , MMTTCC22
1 bit required to indicate power per eigenvector1 bit required to indicate power per eigenvector
42
Numerical ResultsNumerical Results
Illustrating convergence, 2-user BC Illustrating convergence, 2-user BC
0 0.5 1 1.5 2 2.5 3 3.5 40
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Rate Allocated, Power
Prob(Rate > x) or Prob(Power > x)
2x2, PMAX = 10, Rate constraint = [1, 1]
0 100 200 300 400 500 600 700 800 900 10000
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Iteration Count
Rate or Power
2x2, PMAX = 10, Rate Constraint = [1, 1]
single channel realization single channel realization 100 channel realizations 100 channel realizations
43
DiscussionDiscussion
Presented an iterative algorithm that achieves aPresented an iterative algorithm that achieves aspecified rate vector in the capacity region of aspecified rate vector in the capacity region of aMIMO Broadcast ChannelMIMO Broadcast Channel Convergence demonstrated (see CISS Convergence demonstrated (see CISS ‘‘04 paper)04 paper)
Proposed algorithm has a limited feedbackProposed algorithm has a limited feedbackinterpretationinterpretation Incremental transmit covariance updates obtainedIncremental transmit covariance updates obtained
from each receiver using from each receiver using MMTTCC2 2 + M+ MT T bits per userbits per user
44
OutlineOutline Introduction to the WNCG @ UT Austin
Scheduling in MIMO ChannelsScheduling in MIMO Channels
User-Level Performance LimitsUser-Level Performance Limits
Limited FeedbackLimited Feedback
Conclusions
45
ConclusionsConclusions Proposed a model for scheduling in MIMO BCs
Evaluated the performance of various schedulers Max sum rate scheduling yields good delay perf
Note this does not include file length or delay For low SNR values there is a crossing point
Max rate or shortest remaining file size may be better
Proposed iterative feedback algorithm forachieving points in BC rate region Algorithm has a limited feedback interpretation
46
Ongoing WorkOngoing Work Evaluation of other scheduling disciplines Effect of low complexity receivers on performance
Extend our work in [HeaAirPau’01] to include delay Evaluate the performance of other receivers
Transmit beamforming versus DPC Iterative feedback with other quantized feedback