Abstract
Multicast protocols that provide message ordering and delivery guarantees are becoming increasingly important in distributed system design. However, despite the large number of such protocols, little analytical work has been done concerning their performance, especially in the presence of message loss. This paper illustrates a method for determining the performability of group multicast protocols using stochastic activity networks, a stochastic extension to Petri nets, and reduced base model construction. In particular, we study the performability of one such protocol, called Psync, under a wide variety of workload and message loss probabilities. The specific focus is on measuring two quantities, the stabilization time - that is, the time required for messages to arrive at all hosts - and channel utilization. The analysis shows that Psync works well when message transmissions are frequent, but it exhibits extremely long message stabilization times when transmissions are infrequent and message losses occur. We use this information to suggest a modification to Psync that greatly reduces stabilization time in this situation. The results provide useful insights into the behaviour of Psync, as well as serving as a guide for evaluating the performability of other group multicast protocols.