Abstract KMPI: Kernel-Level Message Passing Interface

MPI is one of the most widely used APIs for parallel supercomputing and appears to map well to a large set of problems. This paper describes our in-kernel implementation of MPI that by-passes multiple OS layers and protocols to connect user code directly to networking hardware. Performance analysis shows that this implementation is significantly (2 to 4 times) faster than the TCP/IP based versions and has impressive performance when compared with versions using shared memory. As important, our MPI implementation is much simpler than the reference standard MPICH and LAM versions, and requires less by way of auxiliary support software.