Classification and Analysis of Distributed Event Filtering Algorithms Sven Bittner and Annika Hinze University of Waikato, New Zealand, {s.bittner, hinze}@cs.waikato.ac.nz Abstract. Publish/subscribe middleware provides efficient support for loosely coupled communication in distributed systems. A number of different distributed message-filtering algorithms have been proposed. So far, a systematic comparison and analysis of these filter algorithms is still missing. This paper proposes a classification scheme for distributed filter algorithms that supports the theoretical and practical analysis of these algorithms. We present a first cut theoretical evaluation and a subsequent practical evaluation of promising candidate algorithms. Factors that are considered include the characteristics of the underlying network and application-related constraints. Based on the findings of these evaluations, we conclude with a summary of the strengths and weaknesses of the algorithms that we have studied. 1 Introduction Large scale distributed systems increasingly rely on middleware-level publish/subscribe services to implement loosely coupled communication between components. The ex- changed messages are filtered and forwarded to the appropriate components. This paper proposes a classification of distributed filter algorithms and provides an extensive theo- retical and experimental analysis of selected algorithms. An event notification system or publish/subscribe system is a middleware imple- menting the event-based communication paradigm. A publisher component sends event messages that announce the occurrence of events, i.e., the occurrence of something of interest within the distributed system. Subscriber components can subscribe to events that are of interest to them; these subscriptions are called profiles. Components can act as publishers and/or subscribers. The publish/subscribe system filters the incoming messages according to the subscribers’ profiles and forwards matched messages to the respective subscribers. The distributed components of the publish/subscribe system are referred to as brokers. We now briefly describe the current situation from which we will derive the re- search questions that are addressed in this paper. Several distributed algorithms have been proposed for the efficient filtering of event messages based on the context of the messages [1, 3, 7, 9–12, 14]. Rendezvous nodes [11, 12] are particular brokers that spe- cialize in the filtering of selected event types and act as meeting points for profiles and event messages. Rendezvous nodes are a combination of a centralized and a distributed