Supporting nondeterministic execution in fault-tolerant systems

We present a technique to track nondeterminism resulting from asynchronous events and multithreading in log-based rollback-recovery protocols. This technique relies on using a software counter to compute the number of instructions between nondeterministic events in normal operation. Should a failure occur, the instruction counts are used to force the replay of these events at the same execution points. The execution of the application thus can be replayed to recreate the pre-failure state, while accommodating uncontrolled nondeterminism during normal operation. Implementation on a DEC Alpha processor shows that this support has a low overhead, typically less than 6% increase in running time for the applications we studied.