Estimating cache performance for sequential and data parallel programs

Abstract

This paper introduces an analytical model that enables automatic estimation of the cache performance for both sequential and data parallel Fortran programs. The estimation is based on a classification of array accesses with respect to cache reuse at the source code level. An estimated upper bound of the number of distinct cache lines accessed inside of a loop is statically computed. Based on this estimate the number of cache misses for loops, procedures and the entire program can be predicted.

The method has been implemented as part of P ³ T (Parameter based Performance Prediction Tool) and successfully supports VFCS (Vienna Fortran Compilation System) in guiding the application of data distributions and program transformations on distributed memory multiprocessor systems to achieve greater cache effectiveness. Experiments are presented that demonstrate the efficacy of our approach with very encouraging experimental results.