Basing rollback recovery on optimistic message logging and replay avoids the need for synchronization between processes during failure-free execution. Some previous research has also attempted to reduce the need for synchronization during recovery, but these protocols have suffered from three problems: not eliminating all synchronization during recovery, not minimizing rollback, or providing these properties but requiring large timestamps. This paper makes two contributions: we present a new rollback recovery protocol, based on our previous work, that provides these properties (asynchronous recovery, minimal rollback) whilereducing the timestamp size; and we prove thatno protocol can provide these properties and have asymptotically smaller timestamps.
Sean W. Smith, David B. Johnson