01.27.09
The NASA Advanced Supercomputing (NAS) Division's Application Performance and Productivity (APP) group recently demonstrated that some OpenMP codes scale better and run faster per-processor core on the Schirra supercomputer (an IBM POWER5+) than on Columbia (an SGI Altix)

As a result of some impressive speed-ups obtained on Schirra for one user's code, SPH3D, the APP group wrote a script to regularly identify OpenMP codes running on Columbia, and to determine their parallel computing efficiency. Poorly performing codes were then targeted for porting to Schirra. To date, the OpenMP codes of three users have been transferred to Schirra to improve performance

One such code, named IRFS90_T21, run by a user at NASA Glenn Research Center, has a parallel efficiency of only 20-30% on the SGI Altix 3700s and 30-40% on the Altix 4700s that are part of the Columbia system. To give the researcher a "running start," APP group member Johnny Chang ported the code to Schirra and obtained parallel efficiencies better than 60%. Another code, a global climate modeling code from NASA Goddard Institute for Space Studies (run by a user at Columbia University's Center for Climate Research), was also moved to Schirra and obtained better OpenMP scaling.

In addition to identifying candidate OpenMP codes for migration to Schirra, the data collected by the APP group provides detailed information on which OpenMP codes are running on Columbia and who the users are. This information will also be useful in future procurements to provision resources needed for running OpenMP codes.

The 40-node Schirra supercomputer gave impressive results in speeding up three OpenMP codes ported from Columbia.

For more information about this activity, please contact: Johnny Chang, Johnny.Chang@nasa.gov

For information about NASA and agency programs on the Web, visit http://www.nasa.gov/home/