A publication of the Advanced Simulation & Computing Division, NA-121.2, NNSA Defense Programs
June 2008
NA-ASC-500-08—Issue 7
Return to this issue’s stories
NIF Full-Beam Simulation on BlueGene/L a Success
Lawrence Livermore National Laboratory recently completed a parallel F3D (pF3D) simulation of a National Ignition Facility (NIF) 50-degree laser beam on the BlueGene/L computer. This is the first time that a simulation has modeled the full cross section of a NIF beam from the laser entrance hole to the hohlraum wall. Until now, simulating the entire beam cross section was beyond the reach of available computational resources. The BlueGene/L full-beam simulation provides a benchmark for evaluating the effects of approximating the beam by a central slab.
The simulation ran for about 10 days and 13 hours on 196,608 processors with approximately 24 billion zones, and pF3D demonstrated nearly ideal performance scaling on this massively parallel calculation. “This appears to be the highest processor count ever employed in a multiphysics simulation, and the results are being submitted for consideration for the 2008 Gordon Bell Prize,” said LLNL Physicist Steve Langer. “Most Gordon Bell winners in the past have had a single compute-intensive kernel that consumed over 90% of the run time and had moderate amounts of communication over the interconnect.”
According to Langer, 20 different functions accounted for 95% of the compute time during the pF3D run. Some of these functions perform significant amounts of memory access, so pF3D could not achieve as high a percentage of peak floating-point performance as “single-kernel” codes. A pF3D process communicates extensively with many other processes, not just with nearest neighbors.
“These characteristics of pF3D make it harder to run efficiently at very high process counts than earlier Gordon Bell winners,” Langer explained. “A number of other submissions will achieve a higher teraFLOPS rate than our pF3D run. Our multiphysics simulation running efficiently on very high processor counts is a major step forward.”
The incident laser simulates Brillouin scatter from two distinct regions.
DOE Privacy Disclaimer | Sandia Privacy Disclaimer | 2008-4454 W
Developed and maintained by Sandia National Laboratories for NA-121.2