Accelerator Aware MPI Micro-benchmarking using CUDA, OpenACC and OpenCL

Recently MPI implementations have been extended to support accelerator devices, Intel Many Integrated Core (MIC) and nVidia GPU. This has been accomplished by changes to different levels of the software stacks and MPI implementations. In order to evaluate performance and scalability of accelerator aware MPI libraries, we developed portable micro-benchmarks to indentify factors that influence efficincies of primitive MPI point-to-point and collective operations. These benchmarks have been implemented in OpenACC, CUDA and OpenCL. On the Intel MIC platform, existing MPI benchmarks can be executed with appropriate mapping onto the MIC and CPU cores. Our results demonstrate that the MPI operations are highly sensitive to the memory and I/O bus configurations on the node. The current implemetation of MIC on-node communication interface exhibit additional limitations on the placement of the card and data transfers over the memory bus