Reducing Logging Overhead for Deterministic Execution

Deterministic execution of parallel applications can be enforced by logging speculative memory accesses and restoring saved state in the event of a conflict. However, logging all accesses has a prohibitive overhead. We propose three techniques to reduce logging in an always-on transactional memory system. The first, Threaded Section Analysis, statically determines when only a single thread is running, so instrumentation is unnecessary. The second, Loop Invariant Log Motion, reduces logging requirements for loops by moving logging operations outside loops. The third, Multi-Threaded Read Only Memory, reduces instrumentation significantly using simple programmer inserted directives. When combined, these techniques reduced logging requirements by an average of 63% in the benchmarks we evaluted.