Architectural Support for Reliable Parallel Computing

Abstract

ReEnact improves reliability of multi-threaded software by providing an effective data-race debugging support. ReEnact extends thread-level speculation (TLS) mechanisms to roll back the buggy execution and repeat it as many times as necessary until the bug is fully characterized. These incremental re-executions are deterministic even in multi-threaded codes. The specific implementation of ReEnact detailed and evaluated in this thesis targets data-races in multi-threaded programs. Our experiments using SPLASH-2 applications show that ReEnact is very effective at detecting and characterizing data-race bugs automatically on the fly. This we consider the most valuable contribution of ReEnact. Moreover, in many cases, ReEnact also repairs the bug. Last but not least, ReEnact is fully compatible with production runs: the slowdown of race-free execution is on average only 5.8%.