A methodology for the performance prediction of massively parallel applications

This paper presents a methodology to predict the execution time of massively parallel applications before any significant implementation actions are taken. This methodology captures the problem decomposition into tasks and their precedence relationship, along with the computational and communication demands placed by the application on the underlying architecture. An example shows how the methodology may be used to study the effects of various data placement strategies, problem size, and number of processors for an LU factorization algorithm. The model predictions were validated with published experimental results on a Touchstone Delta machine