This thesis is motivated by the vision of enabling the ubiquitous deployment of applications such as audio/video conferencing and broadcasting over the Internet. For over 15 years, researchers have attempted to enable such applications using the IP Multicast architecture. However, concerns regarding per-group state in routers, deployment issues, and difficulties with supporting higher level functionality such as reliability and congestion control has prevented IP Multicast from taking roots.

We take the stand that "it is feasible to efficiently enable group communication applications on the Internet without router and IP level support". We demonstrate this thesis in the context of an alternate architecture that we call End System Multicast. Here, end systems implement all multicast functionality, including membership management, and packet replication. By eliminating state in routers, and exploiting application-specific intelligence, we argue that End System Multicast can address fundamental concerns with IP Multicast.

We present the design and implementation of protocols for constructing efficient overlays among participating end systems in a self-organizing manner. The scale of nodes involved, and the dynamic and heterogeneous nature of the Internet make the design of these protocols different than traditional distributed algorithms. We present Narada, the first published self-organizing protocol for overlay multicast. We also present Sparta, a protocol deployed in a fully operational broadcasting system based on End System Multicast. The system has been used to broadcast several events including the ACM SIGCOMM and SOSP conferences, and has been used by several thousand users. The thesis adopts an integrated approach to validating architecture, protocol design, and systems building. The protocols address issues such as constructing bandwidth-optimized overlays, and node heterogeneity, that are critical in building operational systems, yet overlooked by the community.

The thesis has influenced the community's thinking on multicast and inspired much follow-on effort. Narada has been extensively used as a benchmark for comparison. Metrics that the thesis introduces, such as Stress, and Relative Delay Penalty, have become standard benchmarks for evaluating overlay based solutions. Experience gathered from extensive real deployment is a distinguishing highlight of this thesis.

137 pages

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by reports@cs.cmu.edu