Optimal sampling-based rate adaptation: implementation and performance evaluation

Abstract

Rate Adaptation (RA) is a fundamental mechanism in 802.11 systems which allows the senders in the wireless networks to adapt the modulation and coding scheme as well as the MIMO transmission mode to the wireless channel conditions. For example, a sender may select a (mode, rate) pair on a per-packet basis from a finite set of MIMO transmission modes and transmission rates. The objective of RA mechanism is to maximize the throughput by identifying the (mode, rate) pair as fast as possible which maximizes the product of the rate and the transmission success probability. The challenges in terms of the design of RA algorithm arise from the facts that (i) the transmission success probabilities are not known in advance at the senders, and (ii) these probabilities may evolve over time. RA algorithms, therefore, should estimate the wireless channel conditions in order to learn and track the (mode, rate) pair providing the highest throughput. Basically, there are two approaches to estimate thewireless channel conditions. The first one is the sampling-based approach, and the second one is the SNR-based approach. It is difficult to predict which approach will be generally adopted in the future, however, the samplingbased RA mechanisms are more popular in the current 802.11 systems.In this thesis, we focus on the sampling-based RA approach. Most sampling-based RA algorithms are driven by heuristics so far. In constrast, we consider a sampling-based RA algorithm based on theory. More specifically, a sampling-based RA algorithm can be formulated as an online stochastic optimization problem, and solved by mapping the problem into a MAB (Multi-Armed Bandit) problem. The sampling-based RA algorithm driven by solving the MAB problem is asymptotically optimal for stationary radio environments. However, in practice, the wireless channels may be non-stationary due to mobility of stations in the wireless networks, hence the samplingbased RA algorithm based on theory cannot be used ...