Two new problems in congestion control: MAC3, and restless bandits
Seminar Room 1, Newton Institute
Abstract1. Medium Access Control (MAC) in a wireless network is meant to manage contention between stations that all want access to the shared wireless medium. I will describe a different approach, Medium Access Coding and Congestion Control (MAC3), inspired by the Zigzag algorithm of Gollakota and Katabi (2008). I will argue that MAC3 should be more efficient than contention-based MAC, and that it will automatically solve the problems of "hidden terminals" and "exposed terminals". (Joint work with Devavrat Shah.)
2. At UCL we are implementing a multipath version of TCP. We found that our naive translation of fluid models into packet-level algorithms gave very unsatisfactory performance, since it did not take account of the stochastic nature of congestion feedback. I propose that we should instead derive congestion control algorithms through dynamic programming, in particular using a restless bandit model. I will demonstrate a modified TCP derived through dynamic programming, and I will discuss open issues.