Service Placement in Stream-Based Overlay Networks

May 5, 2005
Peter Pietzuch | Harvard University

The emergence of data stream applications that collect, process, and deliver real-time stream data to multiple consumers in the Internet has created a need for an overlay network designed specifically to support this new class of applications. It is important that such an overlay leverages Internet resources by supporting in-network processing of streams along the transmission path. A key challenge here is how to select the overlay nodes to route and process data to meet the performance goals of the applications. At the same time, the global impact of potentially thousands of streams on the network must not be ignored, and stream paths must adapt to changing network and node conditions.

In this talk, I describe a Stream-Based Overlay Network (SBON) that allows multiple data stream applications to create efficient data paths through the network. I present a relaxation-based, decentralized and scalable technique for the placement of processing services on SBON nodes. A deployment on PlanetLab shows that this approach provides low latency to applications while minimizing network usage and adapting to dynamic changes.

Speaker Details

Peter Pietzuch is a post-doctoral researcher in the Systems Research Group with the Computer Science Department at Harvard University. His main research interests are large-scale distributed systems, including stream-processing systems, publish/subscribe systems, and peer-to-peer overlay networks. Prior to joining Harvard, he received a Ph.D. degree from the Distributed Systems Research Group at the University of Cambridge Computer Laboratory, working on event-based middleware architectures. In 2000, he obtained a B.A. in Computer Science, also from the University of Cambridge.