Functional Programming Helps Speed up MPICollective

This paper shows how functional programming techniques have been successfully applied to relieve a performance bottleneck in an MPI collective communication operation. A functional speciication for the improved operation is given, and an eecient version implementable in an imperative language is derived and proved correct. Experimental results on an IBM SP2 distributed{memory machine connrm the eecacy of the proposed solution. The Message{Passing Interface MPI94, SOHL + 96] is designed to be a complete deenition of the message{passing programming model. MPI has over 120 diierent operations, ranging from primitive point-to-point communication functions to powerful collective operations. It aims to ensure portability of software without introducing a substantial performance penalty, and has been eeciently implemented on a range of hardware architectures. Parallel functional language implementations can beneet from using MPI as an oo-the-shelf building block. The powerful collective operations in MPI ooer a way of achieving good and portable performance. Furthermore, the MPI specii-cation has features which are well suited to functional languages. In particular, MPI derived data types exploit the potential capability of hardware to communicate non-contiguous data (e.g. heap cells) in a single message directly, without the need for memory transfer. MPI has built-in support for writing libraries, it is thread-safe by design, and caters for heterogeneous networks. Apart from vendor{supplied MPI libraries for speciic platforms, a number of widely supported`generic' implementations exist, most notably MPICH MPI], the Portable MPI Model Implementation by the Argonne National Laboratory and Mississippi State University, and LAM BDV94] of the Ohio Supercom-puter Center. Such generic libraries may reasonably be expected to underperform machine{speciic implementations. As evidence of the performance superiority of a specialised library consider Figure 1. It depicts the wallclock times for Integer Sort (IS) from the NAS Par