Predicting MPI-Based Parallel Application Performance on Workstation Clusters Using LogGP
Michael Brim
Abstract:
As clusters of commodity workstations continue to grow in popularity, so does their use
as high-performance parallel application architectures. Current high-performance
computing (HPC) clusters have processor counts in the hundreds, and many initiatives
are already underway that plan to incorporate the use of clusters with thousands to tens of
thousands of processors. Furthermore, much research is ongoing in the area of
computational grids, where distributed HPC sites can be used together in order to solve
ever-increasing size problems. In planning these large-scale systems, most researchers
have been focusing on the hardware and software infrastructure requirements, while
projections of application performance on these systems are few. The goal of this project
is to develop a model of a common HPC parallel benchmark application that will enable
accurate predictions of the performance of similar applications on these large-scale
systems.
Paper available as: Postscript
or PDF
Project presentation available as: Postscript