Model-based analysis and control in stochastic networks

Gang Wu, Purdue University

Abstract

Network models embody rich network knowledge, discovered by networking researchers over a long period of time. We explore in this dissertation effective methods to utilize knowledge embedded in these models to solve networking problems, and demonstrate that exploiting network models enables both performance analysis of realistic communication networks and the design of control methods to drastically improve network performance. Most previous work on performance analysis does not include random fluctuation in bandwidth available to controlled network traffic sources. In this dissertation, we address this deficiency by explicitly including randomness in bandwidth. We also include non-negligible round-trip times in the problem formulation. We analyze the performance of networks whose traffic is dynamically adjusted by distributed congestion-control methods widely deployed in the current Internet. We show that such controlled networks attain optimal performance asymptotically, and we provide scaling laws for tuning control parameters and for dimensioning buffer sizes for improved transient responses, when the network bandwidth grows large. When the buffer size cannot be dimensioned as required or randomness in available bandwidth is overly adverse, current congestion-control methods in the Internet can be severely sub-optimal, calling for new designs. We develop a framework that exploits traffic models in synthesizing controllers to combat large available-bandwidth variations, through online simulation and trace analysis. Under the framework, we construct several controllers by optimizing some performance metrics. Our experiments indicate that these optimization-based controllers achieve significantly improved performance. Optimization-based control design that exploits network models applies to a wide range of network problems other than congestion control. In this dissertation we identify an important buffer-control problem in proxy-based online video streaming, where we model loss events in a network as a self-exciting point process and design a family of controllers that achieve superior performance. The family of controllers apply to any data transfer involving proxy servers and remotely stored data.

Degree

Ph.D.

Advisors

Givan, Purdue University.

Subject Area

Electrical engineering|Systems science

COinS