Towards Scalable Content-based Publish-Subscribe Networks
Architecture design for content-based publish-subscribe service networks is challenging for two reasons: first, communication in such systems is guided by the content of event publications and subscriptions, rather than by network addresses; second, while an event often matches subscriptions from multiple locations, the highly diversified matching patterns and hence communication needs imply that existing multicast techniques cannot be readily used for efficient event delivery.
In this paper, we propose a new architectural approach called MEDYM - Match Early with DYnamic Multicast. Unlike existing approaches, MEDYM does not build static overlay networks for event delivery. Instead, an event is matched against subscriptions as early as possible, to identify destinations with matching subscriptions; then, a multicast tree is dynamically constructed to route the event to the destinations with high network efficiency.
We evaluate the MEDYM architecture using detailed simulations, and compare it with the two major existing design approaches: Content-based Forwarding and Channelization. Experimental results show that MEDYM significantly improves efficiency of event delivery, and is highly flexible and robust.
We also analyze potential overheads introduced in MEDYM, and found them to be well acceptable and more than outweighed by the benefits of the approach. We expect the basic MEDYM architecture to scale to pub-sub networks of thousands of servers, which we believe is adequate for many interesting applications.