The design of a fast and flexible Internet subscription system using content graphics

Abstract

This dissertation describes the design and evaluation of the Fast, Flexible Forwarding system (F3), a distributed system for disseminating information to networked subscribers. It examines existing subscription approaches, proposes F3 as an alternative to these approaches, and presents results from comparisons of F3 and other subscription approaches. Existing subscription approaches examined in the dissertation fall into three categories: unicast, single-identifier multicast, and content-based multicast systems. Careful examination of these approaches suggests that none is able to support complex subscription requests from large numbers of subscribers at high data rates. F3, the systems proposed as an alternative, shares many features with other multicast systems. Like many multicast systems, for example, F3 uses an overlay network of routers to distribute messages to subscribers. F3 differs from other systems, however, in its use of preprocessors to analyze messages before routing begins. Preprocessors carry out analyses of the relationships between subscription topics, and store the results in special content graph data-structures. Preprocessors share the results of their analyses by distributing content graphs to routers in the F3 network. Using content graphs, F3 routers can determine the relationships between subscriptions and notifications more efficiently than in previous approaches. Four studies compared performance of F3 and competing subscription systems. In the four studies, subscription systems handled such tasks as disseminating baseball scores, distributing traffic alerts, and disseminating generic subscriptions formatted as attribute-value pairs.

(cont.) The four studies examined system performance in both simulated network environments and on a working router. Performance characteristics examined in the studies included size of forwarding tables and processing speeds at routers. Results from these experiments showed that F3 does not overproduce messages, as do unicast systems. F3 also outperformed single-identifier multicast systems in such areas as message production, table size, and subscription overhead. The most significant finding of the studies, however, was that F3 processing speed surpassed the speed of a state-of-the-art content-based system by orders of magnitude in scenarios with large numbers of subscribers. Overall, these results suggest that F3 is a promising development in the area of Internet subscription systems.