The performance of data parallel programs suffers from the latencies caused by the memory access and synchronization. These latencies are affected by the grain size and the scheduling policy for the grains since they influence the memory reference patterns and the time required for synchronization among processors. In this paper, to explore the causes of these latencies, data parallel programs are experimenting with varying their grain sizes under two scheduling policies through execution-driven simulations. From this simulation, we found the best grain size and scheduling policy for each simulated parallel programs and analyze the causes inducing those results
.