Optimizing Task Parallelism with Library-Semantics-Aware Compilation

With the spread of parallel architectures throughout all areas of computing, task-based parallelism is an increasingly commonly employed programming paradigm, due to its ease of use and potential scalability. Since C++11, the ISO C++ language standard library includes support for task parallelism. However, existing research and implementation work in task parallelism relies almost exclusively on runtime systems for achieving performance and scalability. We propose a combined compiler and runtime system approach that is aware of the parallel semantics of the C++11 standard library functions, and therefore capable of statically analyzing and optimizing their implementation, as well as automatically providing scheduling hints to the runtime system. We have implemented this approach in an existing compiler and demonstrate its effectiveness by carrying out an empirical study across 9 taskparallel benchmarks. On a 32-core system, our method is, on average, 11.7 times faster than the best result for Clang and GCC C++11 library implementations, and 4.1 times faster than an OpenMP baseline.