Performance Prediction of Benchmark Programs for Massively Parallel Architectures
Jens Simon, Jens-Michael Wierum · 1996
A performance prediction method is presented, which accurately predicts the runtime of a parallel application using the messagepassing model and the asynchronous task programming paradigm. The introduced model also considers the complete memory hierarchyofa multiprocessor node architecture. We study the optimizied parallel Linpack application as an example benchmark program. The predicted performances are compared with measurements on a PowerPC based distributed memory computer with shared memory nodes. We discuss the quality of the model for various processor numbers and problem sizes.