Using the P3T to guide the parallelization and optimization effort under the Vienna Fortran compilation system

Performance prediction of parallel programs is a key issue for the next generation of parallelizing compilers. The paper reports on experiences with the P³T, an automatic parameter based performance prediction tool, which is based on an analytical performance model. It supports the Vienna Fortran Compilation System (VFCS) in parallelizing and optimizing Fortran programs for distributed memory parallel computers. The P³T automatically computes at compile time a set of parallel program parameters which predict the outcome of three of the most crucial performance aspects of parallel programs: work distribution, communication overhead, and data locality. After analyzing the strengths and [imitations of the performance estimator, experiments are shown that demonstrate the ability of the P³T to successfully guide both programmer and compiler in the search for efficient data distribution strategies and program transformations. It is shown that the P³T detects not only crossover points of the goodness of different distribution, strategies but also undulations (or change in slope) of the performance curve for parallel programs