TSXProf: Profiling Hardware Transactions

Abstract—The availability of commercial hardware transactional memory (TM) systems has not yet been met with a rise in the number of large-scale programs that use memory transactions explicitly. A significant impediment to the use of TM is the lack of tool support, specifically profilers that can identify and explain performance anoma-lies. In this paper, we introduce an end-to-end system that enables low-overhead performance profiling of large-scale transactional programs. We present algorithms and an implementation for Intel’s Haswell processors. With our system, it is possible to record a transactional program’s execution with minimal overhead, and then replay it within a custom profiling tool to identify causes of contention and aborts, down to the granularity of individual memory accesses. Evaluation shows that our algorithms have low overhead, and our tools enable programmers to effectively explain performance anomalies. Keywords-transactional memory; profiling; record and repla