On the scalability of many-to-many reliable multicast over the internet

Abstract

Considering the growing desire to interact with multiple users or processes in the Internet, it is well anticipated that a new, promising, and important class of the future Internet applications will require efficient multi-point communication beyond simple point-to-point or unicast communication. The emerging applications lie broadly between one-to-many dissemination-type applications, e.g., stock quotes or file transfer and many-to-many interaction-type applications, e.g., distributed virtual environments or replicated database update, and many of the applications could benefit from reliable multicast. Most existing reliable multicast protocols have focused on one-to -many multicast because the current set of multicast applications driving deployment typically involve a single sender and also it was implicitly expected that the protocols for one-to-many sessions would work effectively in many-to-many session as well, thereby causing many-to-many reliable multicast to receive little attention. In this dissertation, we present a study on many-to-many reliable multicast. We first examine a fundamental question: which class of reliable multicast protocols is most scalable for many-to-many sessions? By developing a new analytical model that considers spatial packet loss correlation and introduces network bandwidth as a new performance measure and comparing more realistic behaviors of the generic reliable multicast protocols, we demonstrate that a key factor for the scalability of many-to many reliable multicast sessions is reducing exposure to retransmission rather than reducing acknowledgment implosion. We also identify a class of the generic reliable multicast protocols that is simple to implement yet the most efficient approach in the aspects of throughput, network bandwidth, and delay. We next design, implement, and evaluate a scalable many-to-many reliable multicast protocol called Group-Aided Multicast (GAM) which is based on the tree-based approach with...