Compiling data parallel programs to message passing programs for massively parallel MIMD systems

Abstract

The currently dominant message-passing programming paradigm for MIMD systems is difficult to use and error prone. One approach that avoids explicit communication is the data-parallel programming model. This model stands for a single thread of control, global name space, and loosely synchronous parallel computation. It is easy to use and data-parallel programs usually scale very well. Based on the experiences of an existing compilation system for data-parallel Fortran programs it is shown how to design such a compilation system and which optimization techniques are required to make data-parallel programs competitive with their handwritten counterparts using message-passing.<>