SKIFCH: HIGH PERFORMANCE COMMUNICATION INTERFACE

We present highly efficient low level message passing interface named SkifCh. This interface is easily supported by network hardware which can be implemented in FPGA (as is in the case for SKIF-Aurora supercomputer) or in specific microchips. SkifCh interface is intended for use as highly efficient communication layer by applications or higher level communication libraries. For the moment there are several libraries implemented on top of SkifCh: MPI, SHMEM, GASNet, and ARMCI. We also provide a performance comparison for SkifCh versus MPI using the SKIF-Aurora supercomputer.

Keywords

supercomputer, HPC, interconnect, message passing interface, SkifCh.