JMVX: Improving Record-Replay for Managed Languages

Record-Replay (RR) is a useful technique for investigating concurrency related bugs whose appearance are subject to particular thread schedulings. RR logs/records the nondeterministic inputs that influence a program such as thread scheduling, shared memory operations, and external file/socket data. The recording can be replayed, providing the same inputs to replicate the bug. Unfortunately, RR is not good at patch validation. Mutable replay is when a recording generated by the buggy version of a program is replayed on the patched version. Patches typically introduce divergent behavior which often prevents replay entirely. In addition, current RR systems are slow for languages running on a Virtual Machine like Java. They limit parallelism while forcing the Garbage Collector and Just-In-Time compiler to behave eactly the same between the recording and replay. RR systems designed for Java replay mulithreaded code incorrectly or require the use of a customized JVM. To alleviate this, we developed a prototype RR system for Java called JMVX. It operates within Java bytecode, allowing it to tolerate benign divergences --- differences in a program's execution caused by functionally equivalent operations introduced by the garbage collector, just in time compiler, and class loading. JMVX supports multithreaded replay by providing a total ordering of monitor/lock operations via a vector clock. JMVX will serve as a foundation to experiment with partial order implementations of the vector clock and specialized support for thread pools. In addition, we plan to support mutable replay by allowing the replay to interact "live" with a mock version of a resource's snapshot.