INDUSTRY

An NCSA Intel Xeon cluster, in use for the past several months to power the center's high-resolution display wall, will bring some of the most sophisticated simulations of tornado data ever done to TV viewers across the country. The cluster, a 40-node, dual processor array running on Intel's Xeon chip was installed at NCSA last fall and has been in use since last November by selected scientists to visualize and render data on a 40-projector scalable tiled display wall capable of resolution as high as 8,192 x 3,840 pixels. The cluster, a 40-node, dual processor array running on Intel's Xeon chip was installed at NCSA last fall and has been in use since last November by selected scientists to visualize and render data on a 40-projector scalable tiled display wall capable of resolution as high as 8,192 x 3,840 pixels. Beginning last December, a visualization team with NCSA's Experimental Technologies division—Robert Patterson, Stuart Levy, Matt Hall, Alex Betts, Lorne Leonard and Division Director Donna Cox—used the visualization cluster to turn simulation data into high-definition animations used in the TV show "Hunt for the Supertwister." The show, an episode of the PBS series 'NOVA,' will air for the first time 3/30/04 at 7 p.m. Central Standard Time (8 p.m. EST) on PBS stations nationwide (check local listings). The visualization cluster consists of 2.4 GHz Intel Xeon nodes, each with 2 gigabytes (GB) of RAM and about 100 GB of local disk space, plus seven I/O nodes, each with about 290 GB of shared disk space. The NCSA team rendered about 150,000 animated frames from computations and analysis carried out by Matt Gilmore, Glen Romine, and Lee Cronce, members of the storm team led by Robert Wilhelmson, an NCSA senior research scientist and University of Illinois professor of atmospheric science. Lou Wicker, a former member of the group and a researcher at the NOAA National Severe Storms Laboratory in Oklahoma, also contributed to the effort and developed the computer code used by the scientists. The frames were rendered at a resolution of 1,920 x 1,080, the quality of high-definition television (HDTV). About 10 minutes worth of animated visualizations done on the cluster made the final cut for the NOVA show. "An astounding amount of graphic rendering was done on this cluster—the work actually required 970 dual-processor CPU days—that's three and half weeks of round-the-clock computer time spread out over three months," said Rob Pennington, NCSA's interim director. "A key point is that no special graphics cards were used. The work was rendered completely with software using the Intel processors. It shows that Xeon has immense capability for this kind of high-quality graphics work." "Hunt for the Supertwister," produced by Tomas Lucas Productions, Inc., examines researchers' efforts to track severe storms, pinpoint the phenomena that trigger tornadoes, and accurately predict in real-time where killer tornadoes are likely to occur. Scientists know that the strongest tornadoes are generated by supercells—a particular type of rotating thunderstorm. Why some of these storms form tornadoes while others don't is still not well understood. Wilhelmson and his research team simulated a supertwister similar to one that ripped through Manchester, SD, last year at 200 mph. Starting with recorded pre-storm conditions near Manchester, the simulation follows the thunderstorm as it erupts and grows and produces a powerful tornado. The simulation produced 650 billion bytes of data and took about 10 days to process on NCSA's IBM p690 system. Using the Intel Xeon visualization cluster, the NCSA team then translated that data into high-definition animated visualizations of the birth and growth of a tornado. The cluster's performance meant the visualization team could regularly test render their work in full HD resolution and evaluate it with the science team. The cluster also allowed the visualization team and the scientists to view 3D versions of their work on NCSA's passive stereo HD display system, a process that gives the scientists a more complete picture of the simulations, but requires twice the computing power of two-dimensional rendering. "The cluster was a wonderful resource," said Levy. "It would have been far more time-consuming to do the job with a traditional short-term, high-performance computing allocation. We needed to study the simulation results quite intensively, which took many weeks of effort. On the cluster, we could keep much of the derived data online because of its large local disk cache. That meant we didn't need to reload hundreds of gigabytes of data for each new rendering run, and that saved us an enormous amount of time."