Comparative Cross-Platform Performance Results from a Parallelizing SML Compiler

We have developed a compiler for Standard ML which allows instantiation of a xed set of higher order functions with equivalent parallel algorithmic skeletons written in C and MPI. The compiler is intended to be both portable and semi-automatic. Here we discuss the performance of the code generated by the compiler, for ve exemplar programs on six MIMD parallel machines. Results suggest acceptable performance and consistent cross-platform behaviour. 1 Compiler overview We have developed a parallelising compiler 10] for a pure functional subset of Standard ML in which the sole source of parallelism is indicated by speciic higher order functions (HOFs). These are translated into parallel algorithmic skeletonss3] implemented in C linked with an MPI library 8]. The main objective of this work is to build a compiler for which all parallelism is implicit. Our use of algorithmic skeletons allows us to partially achieve this goal. On the one hand the selection and parallel implementation of the higher-order functions is automatic. On the other hand the programmer is constrained to use the particular set of HOFs provided by our compiler. We have focused on two HOFs; map and fold: fun map ] = ] | map f (h::t) = f h::map f t fun foldr b f ] = b | foldr b f (h::t) = f h (foldr b f t) map applies a function f to each element of a list and fold, shown here in left-to-right form, applies a function f \between" elements of a list. These functions ?