Abstract:Memory race recording has been proved to be a hard problem in multithreaded deterministic record-replay. It is important to develop an efficient memory race recording algorithm. However, most of the prior work tries to record all memory conflicts, whether they affect deterministic replay or not, resulting a relatively large memory race log. This paper proposes an innovative synchronization aware point-to-point memory race recorder, called SAMR. SAMR analyzes memory conflicts introduced by synchronization operations and classifies them into harmful synchronization conflicts and harmless synchronization conflicts. Harmless synchronization conflicts are filtered out by identifying synchronization operations when recording, and a reduced memory race log is achieved. At the same time, SAMR reduces hardware overhead by using signatures instead of cache memory. Simulations with splash-2 workloads on 8-core CMP system show that SAMR can achieve small memory race size (~ 2 bytes per thousand memory instructions), good scalability in log size and low bandwidth overhead (< 5%), while not needing too much hardware state (~ 1129 bytes). Keywordsmultiprocessor; deterministic replay; memory race recording; synchronization

A Synchronization Aware Memory Race Recorder