Purdue cluster challenge team creates a stir at supercomputing conference

December 2, 2008

Purdue’s team may not have won the Cluster Challenge at the SuperComputing ’08 Conference in Austin, Texas.

Then again, no other team got to run code on a supercomputer powered by bicycling colleagues of seven-time Tour de France champion Lance Armstrong.

The Purdue team also was the first to break the teraflop barrier—a trillion calculations per second—in the student event at the premier international high performance computing conference.

Cyclists from Mellow Johnny’s Bike Shop in Austin, of which Armstrong is part owner, pedal powered a low-power “green” supercomputer from Massachusetts-based SiCortex, Purdue’s partner in the Cluster Challenge, on Thursday as it ran code prepared by the Purdue team.

The program was a retooled version of an historic mathematical simulation of a dam under stress, a problem first solved 100 years ago by a team of human “computers” who worked on it for three years using slide rules. The SiCortex took about 10 minutes in a demonstration that drew a large crowd.

The Purdue team sponsored by Information Technology at Purdue, the University’s central information technology organization, and the Rosen Center for Advanced Computing, ITaP’s research and discovery arm, used a SiCortex for the Cluster Challenge.

Team member Andy Howard, a senior in electrical and computer engineering technology from West Lafayette, said he and his teammates were able to use the competition compliment of about 1,000 processors in the machine to run most of their jobs simultaneously the bulk of the time. But the simulated data, unlike that used in last year’s Cluster Challenge, wasn’t as suited to a processor-rich machine like the SiCortex.

The Purdue team did finish first in the benchmarking phase of the competition Monday, designed to test the limits of the entries at the outset, ahead of a combined team from Dresden University in Germany and Indiana University, which won the overall competition.

The teams are judged on their benchmarking score, the amount of data they can process with designated applications over three days and their overall presentation, preparation and knowledge.

They have to run on a strict 26-amp power limit, too. But shortly after the competition, with the power limit off, the Purdue team and SiCortex beefed up their machine to its full compliment of more than 1,400 processors and ran the benchmarking software again, hitting slightly over a teraflop.

SiCortex builds supercomputers designed to deliver high performance using large numbers of slower, energy-efficient processors, both in power consumption and the cooling they require. Among other things, the company uses a unique, very fast “interconnect fabric”—the wiring that links its processors for working in concert—offsetting the raw speed disadvantage.

Purdue's team consists of Howard; Alex Younts, a sophomore in computer science from West Lafayette; David King, a senior in electrical and computer engineering technology from Lafayette; Paul Willmann, a senior in computer technology from Carmel; and Ryan Weinschenk, a senior in electrical and computer engineering technology from Noblesville.

The students are enrolled in a high performance computing class taught by Jeffrey Evans, an assistant professor in the Department of Electrical and Computer Engineering Technology. Preston Smith, senior UNIX system administrator for Rosen Center, served as team leader.

For the major portion of the competition, Purdue’s and the other Cluster Challenge teams had to run a battery of scientific applications used for such purposes as examining the seismic waves from earthquakes, pollution from automobiles, air flow over a jet plane, evolution at a genetic level and more.

The applications were announced earlier this year, giving the teams months to work with them in the context of their competition machines. In Purdue’s case the unusual architecture of the SiCortex required some adjustments in the software, especially to take maximum advantage of the computer’s voluminous processor array.

The teams didn’t see the simulated data to be run in the applications until it was downloaded to their machines minutes before the biggest part of the competition began Monday evening. That kicked off what Ricky Kendell, Indiana native and Cluster Challenge co-chairman, called “a 44-hour tenacity test.”

“We give them enough work to take seven or eight days and they get 44 hours to do that,” Kendell said.

Smith said the amount of data is, in fact, generally too much for any team to finish. The goal is to get as far as possible. The Purdue team, working in shifts, kept at that task until late Wednesday afternoon when the competition ended.

“I slept part of the time,” said King, coming off the overnight shift Tuesday, although he allowed that the rock music blaring near constantly from loudspeakers next to the Cluster Challenge area at the conference made sleeping a challenge as well.

“I went to the show floor and I could still hear it in my head,” Younts said.

The students normally didn’t have a lot to do once they started a data processing job.

“If something goes wrong we have to fix it,” Weinschenk said. “But until something goes wrong, it’s mainly just sitting here waiting for something to go wrong. It’s kind of like pulling an all-nighter, but not actually doing anything.”

Still, tired or not, they were ready to head to a party thrown for the teams at the end of the competition.

“Eight hours of sleep in the last 48 hours, that’s not bad,” Howard said.

Writer: Greg Kline, (765) 494-8167, gkline@purdue.edu

Sources:

Originally posted: December 2, 2008