A preference clustering protocol for large-scale multicast applications

Abstract

IP Multicast has enabled a variety of large-scale applications on the Internet which would otherwise bombard the network and the content servers if unicast communication was used. However, the efficiency of multicast is often constrained by preference heterogeneity, where receivers range in their preferences for application data. We examine an approach in which approximately similar preferences are clustered together and transmitted on a limited number of multicast addresses, while consuming bounded total session bandwidth. We present a protocol called Matchmaker that coordinates sources and receivers to perform clustering. The protocol is designed to be scalable, fault tolerant and reliable through the use of decentralized design, soft-state operations and sampling techniques. Our simulation results show that clustering can reduce the amount of superfluous data at the receivers for certain preference distributions. By factoring in application-level semantics into the protocol, it can work with different application requirements and data type characteristics. We discuss how three different applications—stock quote dissemination, distributed network games, and session directory services—can specialize the protocol to perform clustering and achieve better resource utilization.