Delay efficient control policies for wireless networks

Gagan Raj Gupta, Purdue University

Abstract

The last few years have witnessed significant developments in our understanding of how to control multi-hop wireless networks for optimizing performance metrics such as throughput and stability. However, the great open challenge has been to go beyond our understanding of stability and throughput-optimality and analyze network performance for a variety of important metrics and obtain efficient control policies. Analyzing delay performance is extremely difficult due to the complex correlations that arise between the arrival, service, and the queue length processes. In this dissertation, we develop novel techniques for the delay analysis of wireless networks and also design new control policies that are nearly delay-optimal or in some cases, delay optimal. We also study the problem of scheduling network switches (cross-bar constrained), which form the building blocks of most network architectures. Unlike traditional large-deviation or heavy traffic approaches that rely on asymptotics, our results are also shown to be accurate in non-limiting regimes. We generalize the typical notion of bottlenecks and develop novel techniques to analyze and control systems with multiple bottlenecks. In wireless networks and switches, bottlenecks arise due to the presence of interference or resource constraints. We show that these bottlenecks impose fundamental limits on the evacuation time and expected delay of a system. Furthermore, we use the concept of scheduling heavy bottlenecks to design near delay-optimal policies. We show that although delay optimal policies can be designed for some special cases like the clique network and small switches without arrivals, in general, it is provably impossible to find optimal policies for certain delay metrics. This result further emphasizes the difficulty of designing policies that provide good delay performance. We then design scheduling policies for switches that are based on the concept of scheduling “enough” heavy bottlenecks. These policies minimize the “evacuation time” (time until the system is drained) for any given initial configuration while maximizing the system throughput. We develop novel algorithms based on the idea of “post-processing” any given schedule by making minimal changes. This leads to a flexible framework to design throughput-optimal policies which satisfy the goals of computational efficiency, near delay-optimality, or fairness as per the needs of the applications. We develop fundamental bounds on the delay performance of control policies for arbitrary multi-hop wireless network under a generic combinatorial interference model. Simulation results show that the lower bound captures the fundamental properties of the wireless networks (interference, spatial-multiplexing, fading etc.). For several representative topologies, we design a control policy that performs close to the lower bound. For networks with single-hop traffic, we develop a policy whose delay performance is no worse than any stationary randomized scheduler. We also develop an upper bound on the delay performance of this policy and show that it is tighter than the state-of-the-art.

Degree

Ph.D.

Advisors

Krogmeier, Purdue University.

Subject Area

Computer Engineering

COinS