Enforcing Strong Consistency with Semantically View Synchronous Multicast

Replication is a fundamental strategy to obtain highly available services. Among the mechanisms that support replication, view synchronous multicast protocols emerge as a powerful abstraction to encapsulate fundamental problems in replication. Unfortunately, in the presence of a temporarily slow processor or network link the performance of implementations of view synchronous multicast is degraded for the whole group. This happens due to the strong reliability criterion, which forces a potentially large number of messages to be stored, eventually leading to buffer exhaustion and intermittent blocking of the application. This paper proposes a new multicast primitive, Semantically View Synchronous Multicast, that alleviates this problem by selectively weakening reliability constraints while, at the same time, allowing strong consistency to be enforced at a higher level. The usefulness and practical relevance of the new primitive is illustrated using a modified primary-backup replication protoco