Scheduling of wavefront parallelism on scalable shared-memory multiprocessors

Tiling exploits temporal reuse carried by an outer loop of a loop nest to enhance cache locality. Loop skewing is typically required to make tiling legal. This restricts parallelism to wavefronts in the tiled iteration space. For a small number of processors, wavefront parallelism can be efficiently exploited using dynamic self-scheduling with a large tile size. Such a strategy enhances intratile locality, but does not necessarily enhance intertile locality. We show that dynamic self-scheduling performs poorly on scalable shared-memory multiprocessors where smaller tiles are necessary to provide sufficient parallelism-smaller tiles place greater importance on intertile locality. We propose static scheduling strategies which enhance intertile locality for small tiles. Results of experiments on a Convex SPP1000 multiprocessor demonstrate that our strategies outperform dynamic self-scheduling by a factor of up to 2.3 on 30 processors