Repeating history: Execution replay for parallel Haskell programs

Parallel profiling tools, such as ThreadScope for Parallel Haskell, allow programmers to obtain information about the performance of their parallel programs. However, the information they provide is not always sufficiently detailed to precisely pinpoint the cause of some performance problems. Often, this is because the cost of obtaining that information would be prohibitive for a complete program execution. In this paper, we adapt the well-known technique of execution replay to make it possible to simulate a previous run of a program. We ensure that the non-deterministic parallel behaviour of the Parallel Haskell application is properly emulated while its deterministic functionality is unmodified. In this way, we can gather additional data about the behaviour of a parallel program by replaying some parts of it with more detailed profiling information. We exploit this ability to identify performance bottlenecks in a quicksort implementation, and to derive a version that achieves an 82% speedup over a naive parallelisation.