Poster: An MPI Library implementing Direct Communication for Many-Core Based Accelerators

DCFA-MPI is an MPI library implementation for manycore-based clusters, whose compute node consists of Intel MIC (Many Integrated Core) connected to the host via PCI Express with InfiniBand. DCFA-MPI enables direct data transfer between MIC units without the host assist. MPI_Init and MPI_Finalize functions are offloaded to the host side in order to initialize the InfiniBand HCA and inform its PCI-Express address to MIC. MPI communication primitives executed in MIC may transfer data directly to other MICs or hosts by issuing commands to HCA. The implementation is based on the Mellanox InfiniBand HCA and Intel´s Knights Ferry, and compared with the Intel MPI + offload mode. Preliminary results show that DCFA-MPI outperforms the Intel MPI + offload mode by 1.12 ~ 5 times.