NeTS-NR: Counter-Intuitive Behavior in General Networks

California Institute Of Technology, Pasadena CA

Investigators

Abstract

Proposal Number: 0435520 PI: Steven Low Institution: California Institute of Technology Title: Counter-Intuitive Behavior in General Networks Abstract It is well-known that counter-intuitive behavior can arise in a distributed system where agents optimize their own objectives. This project studies the intricate interactions in TCP networks. The emphasis is on general network topologies with multiple links and heterogeneous flows and on rigorous mathematical modeling and analysis. The project develops mathematical theory to understand the behavior of large scale networks, verifies these theoretical insights in simulations and real networks, and exploits them in practical design and operation of networks and protocols. Preliminary results illustrate that interesting and counter-intuitive behavior arises only in a network setting where flows interact through multiple shared links in surprising ways. For example, the researchers have shown that increasing link capacity can reduce the aggregate throughput in a TCP network, and that, contrary to conventional wisdom, a fair bandwidth allocation is not necessarily inefficient. These results demonstrate that the formal approach taken in the project is indeed necessary and rewarding. The results of this project will help us better understand behavior of large-scale networks such as the Internet, and in particular, help characterize paradoxical phenomenon and derive guidelines to avoid harmful behavior in TCP networks. Together with Professor Mung Chiang of Princeton EE, the PI is developing a new course on ``Control and optimization of communication networks'', at the introductory graduate and advanced undergraduate level. The results produced in the project will be directly incorporated into the new course. We will also leverage on, and contribute to, the education and outreach activities of this initiative.