Correctness of automated distribution of sequential programs

Overview & Synopsis

In this paper, we prove that the data-driven parallelization technique, which compiles sequential programs into parallel programs for distributed memory parallel computers, is correct. We define a model based on labeled transition systems, and we prove, from the chosen compilation rules, the confluence of all possible behaviours of the parallel programs we obtain, in spite of asynchronism due to the communications. We also show that this model is powerful enough to prove the correctness of various optimizations of the basic compilation mechanism.