Adaptive OpenMP for large NUMA nodes

Abstract

The advent of multicore processors advocates for a hybrid programming model like MPI+OpenMP. Therefore, OpenMP runtimes require solid performance from a small number of threads (one MPI task per socket, OpenMP inside each socket) to a large number of threads (one MPI task per node, OpenMP inside each node). To tackle this issue, we propose a mechanism to improve performance of thread synchronization with a large spectrum of threads. It relies on a hierarchical tree traversed in a different manner according to the number of threads inside the parallel region. Our approach exposes high performance for thread activation (parallel construct) and thread synchronization (barrier construct). Several papers study hierarchical structures to launch and synchronize OpenMP threads [1, 2]. They tested tree-based approaches to distribute and synchronize threads, but they do not explore mixed hierarchical solutions. © 2012 Springer-Verlag.