Performance Modeling and Analysis of Scientific HPC Applications

Event Sponsor: 
Argonne Leadership Computing Facility Seminar
Start Date: 
Feb 4 2008 (All day)
Building 221, Conference Room A216
Argonne National Laboratory
Scott Parker
Speaker(s) Title: 
National Center for Supercomputing Applications, University of Illinois at Urbana-Champaign

Application performance has long been, and continues to be, a crucial concern for the developers and users of scientific applications. Developing software that fully utilizes the power of modern HPC systems presents a significant challenge. Systems composed of hundreds of thousands of processors require highly developed and efficient expression of algorithmic parallelism. While the increasing architectural complexity of multi-core processors makes achieving a significant fraction of peak performance increasingly difficult.

Performance modeling seeks to describe and predict the performance of an application on a target system. Accurate modeling and prediction of application performance is becoming increasingly important as part of the acquisition process for new HPC systems. It is also useful for machine validation, for the selection of algorithms during development, and for identifying performance bounds to guide application optimization.

This talk will examine some of the issues related to achieving high performance on modern HPC systems, as well as the uses and pitfalls of some of the currently available performance tools. Techniques in performance modeling will be discussed along with initial results of performance modeling efforts for the HOMME (High Order Method Modeling Environment) atmospheric baroclinic instability test case.

Miscellaneous Information: