A methodology for improving parallel programs by adding communications

Abstract

We propose a programming methodology for improving the efficiency of applicative parallel programs. The basic idea comes from the program transformation approach à la Burstall-Darlington, in which “new and more efficient” versions of the programs are derived from “old” ones. We extend that approach in the following two respects:we consider parallel concurrent programs, andwe provide a calculus for ensuring that the derived versions are more efficient. We consider that the programmer provides a first version of his/her program and then he/she discovers, maybe in an incremental way, some “facts” about it. Facts are then tested by the calculus and, if they are “accepted”, they will be incorporated into the given version of the program via a translation algorithm, and a new version is thereby derived.