抄録
Non-blocking communication in the MPI (Message Passing Interface) standard provides us with the functionality to overlap communications with calculations. However, the standards only specify the semantics of the APIs (Application Program Interfaces), and actual implementations only assure the context of communications. In fact, the MPI libraries presently installed on a Compaq AlphaServer ES40, a Hitachi SR8000 and the Earth Simulator can not communicate with calculations. We propose a novel parallel strategy to perform the communication and calculations simultaneously by assigning the communication task into one of the physical processors installed on each node. Consequently, we confirm that the present method hides the communication cost behind the calculation one on a Compaq AlphaServer, a Hitachi SR8000 and the Earth Simulator.
