SILICON GRAPHICS SUCCESSFULLY COMPUTES LARGEST COMPUTATIONAL FLUID DYNAMICS PROBLEM

Silicon Graphics, Inc. announced that, on a CHALLENGEarray 16/320 supercomputing server, it has successfully executed the largest computational fluid dynamics (CFD) problem ever. The company, working with the University of Minnesota and the Army High Performance Computing Research Center (AHPCRC), ran the Grand Challenge class problem during September, 1993.

Silicon Graphics and the Minnesota research team decided to work together when professor Paul Woodward discovered that he could not get the desired resolution on the systems at the University of Minnesota/Army High Performance Computing Research Center, one of the largest supercomputing facilities in the world. Working with the Minnesota team, Silicon Graphics put together a parallel system configuration that had 320 100 MHz MIPS(r) R4400(tm) processors, 28 Gigabytes of main memory and 192 Gigabytes of disk.

"When we introduced the Challenge(tm) server product line in January, we claimed we were going to change the supercomputing paradigm. We configured the CHALLENGEarray, optimized the code and ran the largest CFD problem ever within five weeks, demonstrating that we can deliver on our promises," said Forest Baskett, senior vice president of research and development at Silicon Graphics. "With the Challenge line, Silicon Graphics will continue to provide the supercomputing solutions needed for Grand Challenge problems with cost-effective, general purpose, off-the-shelf hardware."

The CFD problem that was computed on the CHALLENGEarray 16/320 was a compressible turbulence analysis problem that utilized a cube of 10243 mesh. In this particular problem, a cube of fluid is divided into 1024 cells in each direction resulting in a total of one billion computational zones. The computational method of domain decomposition was then applied to the CFD problem and mapped onto the CHALLENGEarray architecture to facilitate an efficient method of calculating the problem. The calculation is scientifically valuable because it will provide greater insight into the fundamental nature of compressible turbulence.

"We were very pleased to learn how little reprogramming of our PPM gas dynamics code was required to extract 4.9 Gigaflops out of a possible peak performance of 16 Gigaflops from the CHALLENGEarray 16/320," said University of Minnesota professor Paul Woodward. "We spent two to three weeks tuning our implementation and were delighted with the result. Our collaboration with Silicon Graphics allowed us to attack a grand challenge turbulence problem whose solution had previously seemed beyond the scope of current computational resources."

To solve this kind of CFD challenge, periodic boundary conditions were applied in which an identical cube was mathematically placed on each side, top and bottom of the cube. Data generated from the calculation can also be used to create hypothetical models of turbulence that attempt to do calibrations quickly by using a model rather than solving real equations. The work Woodward and his team have done in compressible turbulence analysis may be used to be understand astrophysics, jet engine turbulence or the behavior of a wake behind a supersonic aircraft.