A performance comparison of asynchronous atomic broadcast protocols

Atomic broaocast ensures that concurent updates to repl'cated data maintahed by a process group are consistently delivered to all group members despite random commmication delays and la'lures. By simplifying the programming 01 applications that use replicated data, atonic broadcast prov'des basic support lor implementing fault-tolerance in dis1r:buted systems. This paper reports discrete event simulation results tnat compare the performance of lour asynchronous atomic broaacast protocols. We hvestigate five performance hdexes: average delivery time, average stabilily time, average number of physical messages sent per update broadcast, maximum buffer s'ze, and oistrioJtion of processing load among group members. These 'ndexes are measured as a function of group size and update inierarrival l'me, both in ihe absence 01 failures and in tne presence of a single commJnicalion failxe. Our comparison showsthat there is no overall besr protocol. W e identify ihose applicat'on areas where a protoco. dominates lne other protocols and we discuss some protocol design techniques lor achieving good performance.