In a parallel or distributed system, relative performance of different programs varies with system and problem size. A determine factor of this performance variation is communication delay. In this study, the influence of communication on scalability is carefully examined. A model is first introduced to estimate communication overhead. The model then is used in a newly proposal iterative method to predict the scaling property of a program automatically. Experimental testing is performed on a Cray T3E and QSW machine to confirm the methodology. Experimental and analytical results show that the combination of estimation and prediction is feasible and effective. It has real potential in performance prediction and program optimization.

Top