Abstract
The authors present the process-replication protocol of Manetho, a system whose goal is to provide efficient, application-transparent fault tolerance to long-running distributed computations. Manetho uses a novel negative-acknowledgment multicast protocol to enforce the same receipt order of application messages among all replicas of a process. The protocol depends on a combination of antecedence graph maintenance, a form of sender-based message logging, and the fact that the receivers of each multicast execute the same deterministic program. This combination allows the protocol to void the delay in application message delivery that is common in existing negative-acknowledgment multicast protocols, without giving up the advantage of requiring only a small number of control messages.< >