Accelerating conversations for fault-tolerant concurrent software

Abstract

Concurrency adds an additional dimension to the problem of developing fault-tolerant software. One of the fundamental approaches for this purpose is the 'conversation'. However, the need for synchronization at the exit from a conversation has a negative effect on performance. The scheme proposed in this paper accelerates conversations by using a special process-a recovery metaprogram (RMP)-which is executed on a separate layer of abstraction-the recovery layer. The RMP, by having access to the history of all interprocess communications, can allow a process that passes its acceptance test to proceed conditionally. If the conditional or any of the other processes does not pass its acceptance test, the conversation must recover and any interactions of the conditional process must be erased. Also, processes which have not reached their acceptance tests can trigger recovery immediately, without waiting for the other processes to get to their individual acceptance tests. This paper presents an architecture and a set of algorithms to implement these ideas.<>