Explaining Intermittent Concurrent Bugs by Minimizing Scheduling Noise

A noise maker is a tool for testing multi-threaded programs. It seeds shared memory accesses and synchronization events (concurrent events) with conditional context switches and timeouts during runtime, in order to increase the likelihood that a concurrent bug manifests itself. However, an instrumented program with many seeded events may not be useful for debugging; events have been seeded all over the source code and provide almost no information regarding the bug. We argue that for many bug patterns only a few relevant context switches are critical for the bug. Based on the observation that bugs involve only a small set of critical events, we develop a randomized algorithm to reduce the scheduling noise and discover these events related to the bug. To evaluate the effectiveness of our approach, we experiment with debugging of industrial code, known open source code software, and programs representing known concurrent bugs. Our results demonstrate that this simple technique is in many cases very powerful, and significantly helps the user locating and understanding concurrent bugs.