MEASURING SINGLE CORE PERFORMANCE OF THE BLUEGENE/P

Carl W Summers1,  Kalyan Kumaran*2,  Vitali Morozov2,  Priyadarshini Malusare3

Purdue University Calumet1, Mathematics, Statistics, and Computer Science, Hammond, IN 46323
Argonne National Laboratory2, Argonne Leadership Computing Facility, Argonne, 60439
University of Southern California3, Computer Science, Los Angeles, 90089

kumaran@alcf.anl.gov


Abstract

MEASURING SINGLE CORE PERFORMANCE OF THE BLUEGENE/P, C. W. Summers1, K. Kumaran*2, V. Morozov2, P. Malusare3, University of Purdue Calumet1, Department of Mathematics, Statistics, and Computer Science, Hammond, IN 46323, Argonne National Laboratory2, Argonne Leadership Computing Facility, Argonne, IL 60439, University of Southern California, Department of Computer Science, Los Angeles, CA 90089, kumaran@anl.gov

The Argonne Leadership Computing Facility houses the largest Open Science supercomputer in the world - a massively parallel BlueGene/P from IBM. The BG/P is composed of 163,840 processors with a peak computational rate of 556 Teraflops. Many applications have been developed on this relatively new platform and are accomplishing breakthrough science. The focus whilst developing these applications has been on scalability, generally leaving the single core performance to the compiler. We would like to refine our expectations of single core performance using a suite of benchmarking tools. Standard Performance Evaluation Corporation’s CPU2006 is an industry standard benchmarking suite designed to measure processor, memory, and compiler performance. CPU2006 is composed of a range of floating point applications from different scientific domains and in a variety of programming languages. CPU2006 has been run on thousands of architectures and those results published, allowing us to better characterize the performance of the PowerPC core in the BG/P. While CPU2006 measures performance in terms of seconds, a better metric for our purposes is flops per second. Using hardware counters we profiled the applications to put performance measurements in terms of flops as well. This talk will outline the process of building, running, and profiling the applications and the results we obtained.

Download

[Abstract (DOC)]