Dr. Lyle Long, professor of aerospace engineering, says, "Lots of people who are using supercomputer centers are running only 16 processors at a time. They may find it more convenient to use these kinds of clusters and to leave the supercomputers to those who need 200 or more processors.

"There are some very important scientific questions in surface chemistry, supercritical fluids, quantum Monte Carlo, chemical kinetics and aeroacoustics that can be tackled using relatively small-scale molecular dynamics, Monte Carlo, integral methods, or finite differences codes," he adds.

"However, one often needs to run these codes repeatedly to either build up statistical data or vary input parameters. One can very easily use the Message Passing Interface (MPI) paradigm, which is free and can be downloaded from the Internet, combined with the Fortran or C programming languages, to run thousands of different cases in a very short amount of time."

The Penn State research detailed two different illustrative applications in a paper, Self-Scheduling Parallel Methods for Multiple Serial Codes with Application to WOPWOP, Tuesday, Jan. 11, at the 38th Aerospace Science Meeting and Exhibit, in Reno, Nevada. His co-author is Dr. Kenneth S. Brentner, senior research engineer, computational modeling and simulation branch, NASA Langley Research Center, Hampton, Va.

Long ran the examples on Penn State's COCOA, the COst effective COmputing Array, a 50-processor cluster of off-the-shelf PCs connected via fast Ethernet. The system, built to study complex fluid dynamics, has 13 gigabytes of RAM memory and 100 gigabytes of disk space. It cost about $100,000 in 1998 and would cost about half of that today. A supercomputer of similar power would cost about $750,000.

"One of the real benefits of inexpensive machines is that they do not have to be shared with hundreds of other users, and we do not have to wait days in a queuing system," Long notes. "We quite often have to wait hours or days at a supercomputer center just to use a few processors.

In addition, while it is quite difficult to get 50,000 CPU hours at a supercomputer center, the COCOA Beowulf cluster provides more than 400,000 CPU hours per year. Furthermore, processors on parallel supercomputers are usually, at most, twice as fast as these PC processors for large production codes."

In their paper, Long and Brentner describe a self-scheduling version of the WOPWOP aeroacoustics code, a standard tool used to predict helicopter noise.

"The WOPWOP code is a relatively small, computationally efficient code that can compute the noise signal from a helicopter rotor in tens of seconds (in some cases) on a scientific workstation," the authors explain.

"Although this sounds like a small amount of time, it is not unusual to compute the noise at thousands of observer locations on a surface to characterize the noise directivity. For example, if the computation for a single observer takes only 30 seconds but there are 1024 observer locations, the total CPU time required will still be more than 8.5 CPU hours."

The researchers used WOPWOP to compute a time history with 512 points at 400 observer locations using three systems with 48 processors: COCOA, a Cray T3E and an SGI 0rigin 2000. COCOA took 127 seconds, the Cray 177 seconds and the SGI 0rigin 2000 just 95. The authors conclude that the results show that "PC clusters can be very effective computing platforms for problems such as these."

Long and graduate research assistant Anirudh Modi have also used COCOA to study the use of Beowulf clusters to predict the full three-dimensional, unsteady, separated flow around complex ship and helicopter geometries which usually requires immense computing resources.

They used a modified version of the flow solver, Parallel Unstructured Maritime Aerodynamics (PUMA) and NASA's VGRID package to generate unstructured grids in order to maximize the number of cells in the area of interest while minimizing cells in the far field.

COCOA was found to be extremely suitable for the calculations, generating a complete, fast and efficient unstructured grid based flow solution around several complex geometries.

"I see a trend to Beowulf clusters. We're already working on our next cluster. The price just falls every year," adds the Penn State engineer.