Downlink Capacity Evaluation of Cellular Networks with Known Interference Cancellation Howard Huang, Sivarama Venkatesan, and Harish Viswanathan Lucent Technologies Bell Labs Motivation Significant advance on known interference cancellation for MIMO broadcast channels – Natural fit with downlink of a cellular system – Most base stations already equipped with 2~4 antennas – Additional processing at the base station is economically reasonable Asymmetric bandwidth requirement for data traffic can justify channel feedback required for known interference cancellation Goal: How much do we really gain? – Best effort packet data – Delay sensitive streaming applications 6/30/2016 Characterization of rate region using duality used for computations Model Mobile receives signal from a single cell and interference from surrounding cells – Phase coordination across multiple cells in outdoor wide area wireless networks appears impractical – Complexity of computing the gains grows with the number of cells Block fading channel model – Mobile feeds back channel conditions from the desired base station in each frame – Ideal noiseless feedback Performance Metrics – Throughput distribution for packet data – Number of users at fixed rate transmission 6/30/2016 System Model Block Fading Wire-Line Network h (t ) h(t 1) h(t 2) h(t 3) h1 h3 6/30/2016 Each interval has sufficient number of symbols to achieve capacity h2 ykn h'k x n vkn Other-cell interference + AWGN Packet Data Throughput In a cellular system different users are at different distances from the base station – Sum rate is a poor metric for comparing gains – A scheduler is used to arbitrate the resources and guarantees some notion of fairness We will use the proportional fair scheduler K – max log(Ti ) where Ti is the long term average i 1 throughput achieved We will assume the backlogged scenario where each user has infinite amount of data to send – Simplifying assumption – Can still obtain reasonable estimate of the gain 6/30/2016 On-line scheduling algorithm In each frame we assign rate vector R R that maximizes K Ri i 1 Ti where Ti is the moving average of the throughput The rate region R depends on the the transmission strategy – DPC rate region when known interference cancellation is employed – Rate region from beam forming 6/30/2016 We have to solve the weighted rate sum maximization in each frame to determine the throughput Maximum Weighted Rate Sum Using duality K max RR BC ( P ) w R i 1 i i K R max K P: Pi P i 1 R MAC ( P ) w R i 1 i i Using polymatroid structure of the MAC rate region w1 w2 wK i K t t max ( w w )log det I H Q H w log det I H Q H i i 1 K l l l l l l K l 1 l 1 Qi : tr ( Qi ) P i 1 K 1 i 1 6/30/2016 Simple proof of optimal ordering For any set of covariance matrices t t det( I + H Q H H 1 1 1 2Q 2 H 2 ) log det(I + H1t Q1H1 ) log det(I + H1t Q1H1 ) t t det( I + H Q H H t 1 1 1 2Q 2 H 2 ) log det(I + H 2Q 2 H 2 ) log det(I + H t2Q 2 H 2 ) 6/30/2016 Since R1 R2 const independent of the decoding order, we should pick the user with least weight to see the most interference w1 w2 Convex Optimization Algorithm Standard convex optimization techniques can be used to perform the maximization Optimization : max f (x) Axb x : Covariance matrices Linear Constraint : Power Constraint Iterative Algorithm Linear Optimization: x* arg max f (x( n)) x x : Ax=b Line Search : t * arg max f tx(n) (1 t )x* t Update : 6/30/2016 x(n 1) t *x(n) (1 t * )x* Beam Forming Scheme Separately encode each user’s signal with zero-forcing beam forming Rate Region for a subset of users (| S | M ) R ZF (S ) R : Rk log det (I + Hk Qk Hkt ), k S Max weighted rate sum within the subset is weighted waterfilling Computing max weighted rate sum over all subsets of users is very complex even for 4 antennas Approx: First select a subset T of users with the highest individual metrics and implement max weighted rate sum only over this subset of users Complexity depends on the size 6/30/2016 KT of the set T Group ZF Beam Forming for Multiple Receive Antennas Similar to group multi-user detection Covariance matrices are chosen such that multiple streams can be transmitted to each user on separate beams Orthogonality of ZF beam forming preserved only across users – The multiple streams for a given user are not orthogonal 6/30/2016 Similar approximation algorithm as in ZF case for computing maximum weighted rate sum Classic Cellular Model MSC BTS Gateway Hexagonal Layout Uniform User distribution 6/30/2016 Simulation Setup 20 users drawn from this CDF 10000 frames with the proportional fair scheduling 6/30/2016 Performance for Single Receive Antenna Factor of 2 improvement w.r.t simple beam forming at 50% point Optimum selection of users with beam forming reduces the gap significantly 6/30/2016 Performance for Multiple Receive Antennas Harder to bridge the gap GZF technique is sub-optimal even among schemes without DPC 6/30/2016 Optimality in a Large Symmetric System Consider a system with large number of users with identical fading statistics – With high probability there will be a subset of users that are orthogonal with high SNR in each scheduling interval Symmetry implies sum rate maximization in each scheduling interval should be optimal – Sum rate is maximized by transmission to subset that is orthogonal with high SNR – Optimal even when joint coding is allowed since sum rate is maximized by transmission to orthogonal subset 6/30/2016 Fixed Rate Evaluation Model For delay sensitive applications we have to guarantee a fixed rate independent of channel conditions – Assume the same rate requirement for all users Translates to determining the equal rate point on the rate region Goal: Evaluate the CDF of number users that can be supported at a given fixed rate (user locations and channel instances are random) – Optimum known interference cancellation – Known interference cancellation with FCFS order – TDMA 6/30/2016 Equal Rate Point on the DPC Region Unable to establish that for any rate vector weight vector w* such that optimization max w* R* R* R* there exists is the solution to the – Cannot iterate on the weights to determine the equal rate point – R* is indeed unique whenever w is such that wi w j for all i j All points of the rate region may not be achievable without ratesplitting or time-sharing For capacity evaluation we need only an algorithm to test if a rate vector is achievable 6/30/2016 Convex optimization algorithm for achievability Define g ( ) max R RR Given a rate vector R* find * arg min g ( ) R* i 1 : i Then R* is achievable iff g ( * ) * R* 0 6/30/2016 Convex Sets and Separating Hyperplanes Can quickly determine points outside the rate region 6/30/2016 FCFS Algorithm Users arrive in some order with the rate requirement Allocate power to the users assuming entire bandwidth is allocated to each user – Use known interference cancellation to remove the new user from interfering existing users – Existing users are interference to new user The arrival order can be sub-optimal Performance will be better than TDMA because of known interference cancellation 6/30/2016 TDMA Vs FCFS (Single Receive Antenna) 50% gain at the 10% point for 4 transmit antennas Gain is not significant for 1 and 2 transmit antennas 6/30/2016 TDMA Vs FCFS (multiple receive antennas) 6/30/2016 FCFS Vs Optimal Ordering MPF – Minimum Power First 6/30/2016 Summary Duality results were used to determine the maximum gain when using a proportional fair scheduler – Factor of 2 gain relative to TDMA strategy with single beam – Single receive antenna case the beam forming can come close to Known Interference Cancellation Algorithm to determine the fixed rate capacity was proposed – 50% improvement relative to TDMA with single beam – TDMA with multiple beams could potentially narrow this gap – Optimum order is comparable to FCFS at the 10% outage level 6/30/2016 Scenarios where inter-cell coordination becomes feasible should be investigated for potentially larger gains