Predicting the performance of large programs on scalable multicomputers

Abstract

The paper introduces the retargetable program-sensitive (RPS) model which predicts the performance of static, data-independent parallel programs mapped to message-passing multicomputers. It shows that the model accurately predicts the performance of mapped programs by comparing RPS predictions to actual execution times in the Poker parallel programming environment. The paper also previews plans for further verification of the model on the NCube2 and other multicomputers.<>