Synchronization Minimization in a SPMD Execution Model

This paper presents an algorithm for synchronization placement when using a SPMD execution model, where synchronizations are enforced only when there exists a cross-processor data dependence. In this paper, we investigate two scheduling techniques, loop-based and data-based, both of which use a SPMD model. Using scheduling information from previous stages in the compilation process, a new technique to determine potential cross-processor data dependences is presented. Given the minimum number of cross-processor data dependences that must be satisfied, a new optimization is used so as to minimize the number of synchronization points needed to satisfy them. This algorithm has been successfully implemented in an experimental compiler. Initial experimental data show this technique to be very effective, outperforming existing methods.