MPI Support for Multi-core Architectures: Optimized Shared Memory Collectives

With local core counts on the rise, taking advantage of shared-memory to optimize collective operations can improve performance. We study several on-host shared memory optimized algorithms for MPI_Bcast, MPI_Reduce, and MPI_Allreduce, using tree-based, and reduce-scatter algorithms. For small data operations with relatively large synchronization costs fan-in/fan-out algorithms generally perform best. For large messages data manipulation constitute the largest cost and reduce-scatter algorithms are best for reductions. These optimization improve performance by up to a factor of three. Memory and cache sharing effect require deliberate process layout and careful radix selection for tree-based methods.