Community Clustering for Distributed Publish/Subscribe Systems

Optimized placement of clients in a distributed publish/subscribe system is an important technique to improve overall system efficiency. Current methods, like interest clustering or publisher placement, treat a client as, either a pure publisher, or subscriber, but not as both. Also, the cost of client movement is usually ignored. However, many applications based on publish/subscribe systems model clients as publisher and subscriber at the same time, which breaks the assumptions made by current approaches. Considering the complex dependency among clients, we propose a new\textit{community-oriented} clustering approach, based on the forming of client clusters that exhibit intense communication relationships, while keeping client movement cost low. The evaluation based on a public data set shows that our method is efficient, adapts to different settings of experimental conditions, and wins over the popular interest clustering approach with respect to number of messages sent, propagation hop count and end-to-end latency.