Abstract

Since the emergence of MPI (Message Passing Interface), how to improve its performance has always been a goal for MPI library implementation. This paper proposes an efficient architecture to achieve this goal from a very low level: using multi-threaded model to implement MPI pointto-point operations in order to overlap communication and computation. Compared to single-threaded active polling communication used by most of the implementations, multithreaded provides more parallelism. A dedicated thread, which is responsible for communication, can spare substantive CPU time for computation. Multi-threaded point-topoint operations have been implemented and tested, and the result is compared with single-threaded implementations in this paper. Both theoretical analysis and experiment result show that the multi-threaded model greatly enhances the efficiency of MPI point-to-point communication. This model is adopted in an ongoing MPI implementation project called FiTMPI.

Citations