Parallel speedup model

A model for speedup of parallel programs

This paper has not been published and should not be distributed.

Abstract

We propose a new model for parallel speedup that is based on two parameters, the average parallelism of the application and the variance in parallelism. We present a way to use this model to estimate these parameters for real applications, using only their observed speedup curves (as opposed to the more complete application knowledge previously required). We apply this method to speedup curves from real applications on a variety of architectures and show that the model fits the observed data well. We propose several applications for these estimated paramters, including the selection of cluster sizes for parallel applications.

Introduction

Conclusions

downey@sdsc.edu