Probabilistic Replay of Java Programs on Multiprocessors

Abstract. Alongside the rise of multiprocessor machines, the concurrent programming model has grown to near ubiquity, being the only possibility for developers wanting to take full advan-tage of their performance. Unfortunately, reasoning about concurrent programs is hard and bugs that stem from unanticipated interactions between tasks are easy to hatch. On the contrary, finding and fixing these bugs is a complex and time-consuming endeavour. To make matters worse, the debugging tools that programmers have relied on for decades fall short when applied to concurrent programs. Systems capable of replaying non-deterministic executions have long been proposed as a solution for this limitation of conventional methods. We survey previous work on deterministic replay systems and present a taxonomy for classifying them. Moreover, we describe the objectives of our future work and how they are to be achieved by proposing an architecture and evaluation methodology for a replayer of concurrent Java programs. The key property of the system is its probabilistic approach: partial production run information is traced and an offline exploration of untraced non-deterministic space is used to achieve deterministic replay. This approach enables an efficient recording phase, at the cost of a potentially longer replay phase which, being done offline, is not as performance critical.