Automatic Parallelization Tool Efficiency: FPP/Atexpert Case Studies

This paper presents several synthetic Fortran codes for testing the ability of automatic parallelizers to generate efficient parallel code. The codes typify sequential solution techniques currently in use at NAS and represent the types of algorithms that automatic parallelizers will be required to analyze and decompose. The Cray automatic parallelizer FPP performed well on codes containing explicitly parallel algorithms, but required user intervention to assist in vectorized relaxation schemes. The inability of FPP to expose parallelism at a level higher than loop level reduced the parallel performance of relaxation and ADI schemes. The Cray utility atexpert consistently diagnosed the causes for poorer-than-expected parallel performance. The current limits on technology indicate that future high speed computer systems will employ multiple processors, operating in parallel, to solve computationally-intensive problems. Efficien