High Performance MPI: Extending the Message Passing Interface for Higher Performance and Higher Predictability

The Message Passing Interface (MPI-1 and MPI-2) provides portability, and a degree of high performance in message passing. A careful review of the design shows many opportunities for performance have been missed, and that these omissions can be corrected. Especially relevant are mechanisms for taking advantage of program characteristics { common to many regular data-parallel and coarse-grain dataow utilizations of MPI { that allow for better runtime optimizability. This paper describes classes of de ciencies and their solutions, through upward compatible strategies that allow \Fast MPI" implementations to exceed the performance of regular MPI implementations on the same platform. A simple approach to retaining portability is also o ered. The most important contribution is the utilization of message passing \temporal locality" in the speci cation of middleware behavior and services between the application and the middleware.