Efficient algorithms for crash recovery in distributed systems

Abstract

We consider the problem of recovering from processor failures efficiently in distributed systems. Each message received is logged in volatile storage when it is processed. At irregular intervals, each processor independently saves the contents of its volatile storage in stable storage. By appending only O(1) extra information to each message, we show that for recovery in general networks O(n2) messages are sufficient and in ring networks ©(n) messages are necessary and sufficient when an arbitrary number of processors fail. By appending O(n) extra information to each message that is sent, we show that O(kn) messages are sufficient for rollingback all of the processors to the maximum consistent states when there are k failures.