Scheduling algorithms for throughput maximization in data networks

Abstract

This thesis considers the performance implications of throughput optimal scheduling in physically and computationally constrained data networks. We study optical networks, packet switches, and wireless networks, each of which has an assortment of features and constraints that challenge the design decisions of network architects. In this work, each of these network settings are subsumed under a canonical model and scheduling framework. Tools of queueing analysis are used to evaluate network throughput properties, and demonstrate throughput optimality of scheduling and routing algorithms under stochastic traffic. Techniques of graph theory are used to study network topologies having desirable throughput properties. Combinatorial algorithms are proposed for efficient resource allocation. In the optical network setting, the key enabling technology is wavelength division multiplexing (WDM), which allows each optical fiber link to simultaneously carry a large number of independent data streams at high rate. To take advantage of this high data processing potential, engineers and physicists have developed numerous technologies, including wavelength converters, optical switches, and tunable transceivers.

(cont.) While the functionality provided by these devices is of great importance in capitalizing upon the WDM resources, a major challenge exists in determining how to configure these devices to operate efficiently under time-varying data traffic. In the WDM setting, we make two main contributions. First, we develop throughput optimal joint WDM reconfiguration and electronic-layer routing algorithms, based on maxweight scheduling. To mitigate the service disruption associated with WDM reconfiguration, our algorithms make decisions at frame intervals. Second, we develop analytic tools to quantify the maximum throughput achievable in general network settings. Our approach is to characterize several geometric features of the maximum region of arrival rates that can be supported in the network. In the packet switch setting, we observe through numerical simulation the attractive throughput properties of a simple maximal weight scheduler. Subsequently, we consider small switches, and analytically demonstrate the attractive throughput properties achievable using maximal weight scheduling. We demonstrate that such throughput properties may not be sustained in larger switches.

(cont.) In the wireless network setting, mesh networking is a promising technology for achieving connectivity in local and metropolitan area networks. Wireless access points and base stations adhering to the IEEE 802.11 wireless networking standard can be bought off the shelf at little cost, and can be configured to access the Internet in minutes. With ubiquitous low-cost Internet access perceived to be of tremendous societal value, such technology is naturally garnering strong interest. Enabling such wireless technology is thus of great importance. An important challenge in enabling mesh networks, and many other wireless network applications, results from the fact that wireless transmission is achieved by broadcasting signals through the air, which has the potential for interfering with other parts of the network. Furthermore, the scarcity of wireless transmission resources implies that link activation and packet routing should be effected using simple distributed algorithms. We make three main contributions in the wireless setting. First, we determine graph classes under which simple, distributed, maximal weight schedulers achieve throughput optimality.

(cont.) Second, we use this acquired knowledge of graph classes to develop combinatorial algorithms, based on matroids, for allocating channels to wireless links, such that each channel can achieve maximum throughput using simple distributed schedulers. Third, we determine new conditions under which distributed algorithms for joint link activation and routing achieve throughput optimality.