MPI and High Productivity Programming

Abstract

MPI has often been called the “assembly language” of parallel programming. In fact, MPI has succeeded because, like other successful but low-level programming models, it provides support for both performance programming and for “programming in the large” – building support tools, such as software libraries, for large-scale applications. Nevertheless, MPI programming can be challenging, particularly if approached as a replacement for shared-memory style load/store programming. By looking at some representative programming tasks, this talk looks at ways to improve the productivity of parallel programmers by identifying the key communities and their needs, the strengths and weaknesses of the MPI programming model and the implementations of MPI, and opportunities for improving productivity both through the use of tools that leverage MPI and through extensions of MPI.