Non-Intrusive Detection of Synchronization Errors Using Execution Replay

Abstract

This paper presents a practical solution for detecting synchronization errors in parallel programs. These errors are: a lack of synchronization resulting in data races, conflicting synchronization resulting in deadlock and redundant synchronization resulting in a performance penalty. The solution consists of a combination of RecPlay, an efficient execution replay mechanism combined with automatic on-the-fly data race detection, deadlock detection and the detection of redundant synchronization during a replayed execution. The detection of data races, deadlocks and redundant synchronization normally introduces an important overhead during an execution, possibly altering the execution. However, by performing these extensive operations during a replayed and therefore unaltered execution there is almost no probe effect. Furthermore, the memory consumption during the data race detection is limited through the use of multilevel bitmaps and snooped matrix clocks. As the record phase of RecPlay is highly efficient, there is no need to switch it off, hereby eliminating the possibility of Heisenbugs because tracing can be left on all the time.