Automatic Parallelization of the AVL FIRE Benchmark for a Distributed-Memory System

Computational uid dynamics (CFD) is a Grand Challenge discipline whose typical application areas, like aerospace and automotive engineering, often require enormous amount of computations. Parallel processing ooers very high performance potential, but irregular problems like CFD have proven diicult to map onto parallel machines. In such codes, access patterns to major data arrays are dependent on some run-time data, therefore runtime preprocessing must be applied on critical code segments. So, automatic parallelization of irregular codes is a challenging problem. In this paper we describe parallelizing techniques we have developed for processing irregular codes that include irregularly distributed data structures. These techniques have been fully implemented within the Vienna Fortran Compilation System. We have examined the AVL FIRE benchmark solver GCCG, to evaluate the innuence of diierent kinds of data distributions on parallel-program execution time. Experiments were performed using the Tjunc dataset on the iPSC/860.