Intelligent Thread-level Speculation

Jeremy Singer

Thread-level speculation (TLS) is a mechanism for improving the execution of sequential programs on multi-core processors. TLS relies on the insertion of potential thread-spawn points into sequential programs, either statically or dynamically. Spawned threads generate parallelism, but they may have to be cancelled during execution if they cause data dependence violations with other concurrent threads.

Until now, spawn points have been identified via (a) exhaustive profiling of programs and/or (b) heuristics devised by human experts. We have recently begun a project that uses machine learning techniques to generate spawning heuristics automatically, with less program profiling. This presentation reports on the use of Java programs' static attributes and dynamic behaviour as features for learning effective TLS spawning policies.