High Throughput Intra-Node MPI Communication with Open-MX

摘要: The increasing number of cores per node in high-performance computing requires an efficient intra-node MPI communication subsystem. Most existing implementations rely on two copies across a shared memory-mapped file.Open-MX offers single-copy mechanism that is tightly integrated its regular stack, making it transparently available to the MX backend many layers. We describe this implementation and offloaded copy using I/OAT hardware. Memory pinning requirements are then discussed, overlapped introduced enable start Open-MX data transfer earlier.Performance evaluation shows local stack performs better than MPICH2 Open-MPI for large messages, reaching up 70% throughput micro-benchmarks when offload. Thanks being involved, also does not heavily depend cache sharing between processing cores, these performance improvements easier observe real applications.