Abstract: Deterministic replay plays an important role in debugging parallel programs. In parallel program design, recording share memory access has become a light spot, according to great non-determinism brought by multithreading memory interleaving. Previous research focused on fine grained instrumentation of memory access instruction, or kernel modification, or special hardware extension to record shared memory communication. However, these methods have the problems of runtime overhead cost and awful potential safety hazard. This paper proposes an efficient deterministic replay technique employing page-protection mechanism on commodity hardware platforms and operating system. Furthermore, this work deeply analyzes the runtime overhead incurred by the proposed technique, and puts forward several optimization strategies of private-lock mechanism, private heap memory pool and initiative page-preemption algorithm to promote the performance. Our prototype UPLAY is implemented on Linux. Performance evaluation shows that the record slowdown is only 9.26X and is much smaller than prior user-level replay system.

Key words: computer software, parallel program, deterministic replay, page protection