ProteusTM: Abstraction Meets Performance in Transactional Memory

International audienceThe Transactional Memory (TM) paradigm promises togreatly simplify the development of concurrent applications.This led, over the years, to the creation of a plethora of TMimplementations delivering wide ranges of performance acrossworkloads. Yet, no universal TM implementation fits each andevery workload. In fact, the best TM in a given workload canreveal to be disastrous for another one. This forces developersto face the complex task of tuning TM implementations, whichsignificantly hampers the wide adoption of TMs.In this paper, we address the challenge of automaticallyidentifying the best TM implementation for a given workload.Our proposed system, ProteusTM, hides behind the TM interfacea large library of implementations. Under the hood, itleverages an innovative, multi-dimensional online optimizationscheme, combining two popular machine learning techniques:Collaborative Filtering and Bayesian Optimization.We integrated ProteusTM in GCC and demonstrated itsability to switch TM implementations and adapt several configurationparameters (e.g., number of threads). We extensivelyevaluated ProteusTM, obtaining average performance < 3%from optimal, and gains up to 100 over static alternatives