PAMcast : programmable any-multicast for scalable message delivery

The rapid growth of the Internet has led to much interest in communication paradigms that support scalability, including multicast and anycast. In this paper, we propose a new message delivery service - Programmable Any-Multicast (PAMcast) - which generalizes both anycast and multicast, by providing a message delivery service to m out of the total n group members, where 1 < m < n. Such a service has potential applications in a number of areas, including fault tolerant repositories, parallel cache queries, and parallel server downloads. Our architecture is based on a shared tree and managed by a group management protocol with similarities to IGMP. To explore what is possible with limited state and computation at the routers, we consider the implementation of two specific modes of delivery that control how the m receivers are selected. Our balanced mode aims for equal distribution of messages over receivers over time. We find that with relatively little state and computation we can achieve this goal. Our closest mode aims to deliver a packet to the m group members closest (by hop count) to the root of the distribution tree. We find that with modest state and computation we can achieve a probabilistic form of the closest mode.