• Summary
• Related Documents
  • Active Bibliography
  • Co-citation
• Version History

Abstract:

Abstract. The deployment of the Multicast Backbone or MBone has enabled a variety of large-scale applications on the Internet. These applications would otherwise bombard the network and the content servers if unicast communication is used. However, the efficiency of using multicast communication is often constrained by preference heterogeneity, where receivers range in their preferences for application data. One approach to accommodate this heterogeneity is to have the sources transmit all their data to the receivers, and the receivers filter out the undesired data. Although this approach is simple to implement, the drawback is that network resources and CPU cycles are wasted in handling the unnecessary data. An alternate approach is to have the sources divide their data into granularity of preferences, and send each preference on a separate multicast address. While receiver preferences are well-matched in this case, this approach can introduce an unacceptable amount of router state and control overhead from using too many multicast addresses concurrently. In this paper, 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. The goal is to maximize preference overlap and minimize network resource consumption. We present a protocol architecture called Matchmaker that coordinates sources and receivers to perform clustering. The architecture is aimed to be scalable, fault tolerant and reliable through the use of decentralized design, soft-state operations and sampling techniques. 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 quotes dissemination, distributed network games, and session directory services---can specialize the protocol to perform clustering and achieve better resources utilization. 1

Citations