Mining Static Features for Squash Prediction in Thread Level Speculation

Thread-Level Speculation (TLS) is a useful technique that facilitates automatic parallelization via optimistic execution of potentially independent threads. It is particularly valuable where there exist ambiguous dependences that cannot be resolved statically by the compiler. However, TLS bears significant overheads due to thread management which can severely degrade any performance gains when the number of squashed threads is high. This study shows the potential of applying Machine Learning techniques in order to construct models that are able to predict whether a thread will commit successfully or otherwise squash. We apply trace-driven simulation for speculative execution using a set of DaCapo Java benchmarks. Our models are able to predict with up to 92% accuracy whether a thread will commit or squash.