Abstract. Atomic broadcast ensures that concurrent updates to replicated data maintained by a process group are consistently delivered to all group members despite random communication delays and failures. By simplifying the programming of applications that use replicated data, atomic broadcast provides basic support for implementing fault-tolerance in distributed systems. This paper reports discrete event simulation results that compare the performance of four asynchronous atomic broadcast protocols. We investigate five performance indexes: average delivery time, average stability time, average number of physical messages sent per update broadcast, maximum buffer size, and distribution of processing load among group members. These indexes are measured as a function of group size and update interarrival time, both in the absence of failures and in the presence of a single communication failure. Our comparison shows that there is no overall best protocol. We identify those application areas where a protocol dominates the other protocols and we discuss some protocol design techniques for achieving good performance.

Print publication: Issue 4 (June 1994)