Identifying pitfalls in automatic parallelization of NAS parallel benchmarks

Abstract

This paper provides an examination of OpenMP based auto-parallelizers and their limitations encountered during parallelization of NAS parallel benchmarks. It also elucidates the issues faced by the parallelizers during parallelization and the resolutions to overcome the problems. Compute-intensive loops are pinpointed using Gprof and the problematic loops within the hotspot area were recognized. Our work concentrates on identifying the pitfalls within the located hotspots and rendering solution in such cases. Analysis on measured speedup and its reasons are well illustrated. This paper underlines the need of a user-interactive environment that highlights the problems evoked during parallelization. It also underscores the obligation for minimal manual intervention concerning coding changes to resolve the problematic code section and make them amenable to parallelization.