A Distributed Implementation method for Parallel Programming

Abstract

A method is described for implementing on a finite network of processing "cells", called the "implementation graph", programs whose potential parallelism is not fixed by the implementation but varies according to the input parameters. First, programming constructs are described permitting a computation, regarded as a dynamic structure called the "computation graph", to diffuse through the implementation graph. Second, the implementation problem of mapping an unbounded number of computation nodes on a finite number of cells is tackled. Processor allocation and message buffering completely disappear from the programmer's concerns. The mechanism proposed is considered a generalization of the stack mechanism.