New directions in optimization-based congestion control

This thesis presents several extensions and improvements to a technique called ‘optimization-based congestion control’ (OBCC). OBCC makes use of a mathematical framework borrowed from microeconomic theory, in which congestion controlled sessions are modelled as self-interested users attempting to maximize their individual utility. The network objective of fair resource allocation is expressed as the maximization of the total utility of all users. To this end, the network provides congestion signals, which are interpreted by sessions as prices for resources to be allocated. When the prices are set correctly, the network objective is maximized as a net result of the sessions making independent self-interested rate-adaptation decisions. The frame-work has been successfully applied to decentralized congestion control for best-effort unicast traffic. Our contributions are to expand the applicability of the technique to other types of traffic and to improve the deployability of practical OBCC protocols in the Internet. Specifically, (1) We extend OBCC to the regime of multicast congestion control, with particular attention paid to the issue of fairness between multicast and unicast sessions. (2) We extend OBCC to a network offering multiple classes of service with the goal of supporting delay-sensitive traffic. (3) Implementing OBCC protocols in the Internet is complicated by the fact that only a single bit in the packet header is available for carrying congestion signals along the forward data path. We address the problem of accumulating the aggregate congestion price along a path under this design constraint, obtaining a complete characterization of possible single-bit protocols and introducing a novel protocol with desirable properties. We also investigate the fundamental limitations imposed by having only a single bit per packet and identify conditions under which stable protocols can be constructed that accurately track the optimal bandwidth allocation