Multicast flow controls in wide -area networks.

Abstract

Multicasting has emerged as one of the most active research areas in the field of networking. This dissertation develops protocols and modeling techniques to solve the new flow-control problems associated with multicast in wide-area networks, such as the Internet. Specifically, it addresses two fundamental components of multicast flow control: (1) <italic>rate control</italic>---adapting the source rate to the available bandwidth; and (2) <italic>flow-control signaling</italic>---conveying the congestion/control information between the source and network/receivers. To make the multicast flow control scalable to round-trip time (RTT) variations, we develop an optimal second-order rate control algorithm, called the alpha-control, to adapt the rate ramp-up speed to the RTT variations in a multicast tree. By modeling the proposed algorithm using the fluid analysis for the multicast ABR (Available Bit Rate) service, we analytically show that the proposed algorithm is stable and efficient, and the resultant source rate and queue size converge to the designated operating regime. The simulation results confirm the analytical findings. Using the <italic>duality</italic> principle, we also model multicast rate control as a distributed optimization problem with a separable structure in aggregate bandwidth utilities and constraints. We then realize the optimization by developing a distributed gradient projection algorithm and a <italic>virtual M-ary</italic> (VMARY) feedback signaling protocol. Our proposed scheme performs as well as any explicit-rate <italic>M-ary</italic> feedback-based scheme while only employing binary feedback. For multicast flow-control signaling, we develop the <italic>Soft Synchronization Protocol</italic> (SSP) to overcome the feedback implosion and noise problems, which consolidates the feedback RM (Resource Management) cells that are not necessarily responses to the same forward RM cell. We also develop the binary-tree models to characterize the delay performance of any given multicast tree. Finally, we develop two stochastic models to capture the statistical delay properties for multicast signaling. The first model focuses on the cases used in RED (Random Early Detection) and REM (Random Early Marking) with independent markings. The second model addresses the general cases with dependent markings. We develop the Markov-chain and Markov-chain dependency-degree models to derive various delay statistics of multicast signaling, and also show that this Markov-chain is ergodic when it is irreducible.