Bill Barth explains how a 16-processor Beowulf cluster serves as a platform for running scientific test cases. Larger problems are placed on CRAY T3E supercomputers at UT and NASA.

With many users sharing resources, access to the largest supercomputers is limited. To sensibly plan their supercomputing sessions, University of Texas at Austin (UT) microgravity researchers first run simple test cases on smaller machines. "We need initial 'validation' runs to ensure the simulation experiment is correctly set up and to compare with repository test cases for debugging the code as it evolves," explained Gurcan Bicken, postdoctoral fellow in the Computational Fluid Dynamics Laboratory.

Tests are a key part of what Laboratory director Graham Carey calls a "hierarchical model of computing." His NASA HPCC team computes most tests on a homegrown Beowulf cluster of 16 Intel Pentium II-based computers [see "Beowulf lives on - as a build-it-yourself computer," INSIGHTS, November 1998]. "Each of these processors has the same amount of memory as a single CRAY T3E processor but only gets about one-half to one-third the performance," said graduate student Bill Barth, "so we can work about the same size problem per processor on this machine. It just takes a little longer."

Then, intermediate-scale scientific studies are carried out on UT's 64-processor CRAY T3E. The biggest simulations are saved for NASA HPCC's 512-processor CRAY T3E at Goddard Space Flight Center and occasionally for larger machines arranged by Silicon Graphics, Inc.

This strategy works primarily because of the way the team's MGFLO (MicroGravity FLOw) software is designed. "The code involves a very modular structure and a high-level preprocessor to enhance portability across different computers," said Robert McLay, Laboratory manager. Moreover, "we spent a great deal of effort in getting great performance per processor and minimizing the cost of adding more processors." With scaling efficiency exceeding 95 percent, MGFLO sustains over 100 billion floating-point operations per second on 1,024 CRAY T3E processors.

Contemporary parallel computers permit high-resolution simulations but present a challenge in managing the reams of data they generate. Files often are too big for most wide-area networks and local workstations to handle, making analysis of results difficult. To get around this obstacle, Laboratory research associate Atanas Pehlivanov is developing an adaptive hierarchical visualization tool that "shifts the hard work to the parallel machine and leaves the smaller tasks to the visualization computer."

During a simulation run, the tool "significantly reduces the data sets selectively, and then we send a much smaller file to the local site and visualize this preview of the results," Pehlivanov said. If interesting physical activity seems to be occurring in a particular region, the user then requests higher resolution in the area and repeats the hierarchic visualization step at this revised resolution level.

Citing plans to release the hierarchical visualization tool to other researchers, Pehlivanov sees the potential for changing how computing is done. "Your environment will be qualitatively different in the sense that once you're capable of quickly seeing your results, you can intelligently guide your simulation session and eventually get better results," he said.