A Novel Approach to Supporting Communicators for In-Switch Processing of MPI Collectives
Abstract: MPI collective operations can often be performance killers in HPC applications, and we seek to solve this bottleneck by offloading them to switch hardware. We've seen from previous works including our own that moving collectives into the network offers significant performance benefits. However, there has been little advancement in providing support for sub-communicator collectives in an efficient manner. Using FPGAs, which provide the ability to couple communication and computation, we have designed an in-switch hardware accelerator to implement support for MPI communicators and full offload of MPI collectives. With our design, preliminary results show that we can achieve up to a 10x speedup over conventional clusters for short message collectives over irregular intra-communicators.
Back to Workshop on Exascale MPI (ExaMPI) Archive Listing
Back to Full Workshop Archive Listing