Abstract

We present a skeleton-based language which leads to portable and cost-predictable implementations on MIMD computers. The compilation process is described as a series of program transformations. We focus in this paper on the step concerning the distribution choice. The problem of automatic mapping of input vectors onto processors is addressed using symbolic cost evaluation. Source language restrictions are crucial since they permit to use powerful techniques on polytope volume computations to evaluate costs precisely. The approach can be seen as a cross-fertilization between techniques developed within the FORTRAN parallelization and skeleton communities.

Citations