ECE Superstars win at supercomputers

The Open Compute team used the most unconventional hardware in their supercomputer, winning them first place at the recent cluster challenge. Photo via David Yates.

The Open Com­pute team used the most uncon­ven­tional hard­ware in their super­com­puter, win­ning them first place at the recent cluster chal­lenge. Photo via Dave Kaeli.

When pro­fessor Dave Kaeli approached them about entering a super­com­puter com­pe­ti­tion, elec­trical and com­puter engi­neering stu­dents Neel Shah and Tushar Swamy had zero expe­ri­ence in the task at hand. “I had never even built a reg­ular com­puter,” said Shah, let alone a super one. But that didn’t stop them from signing up…nor from stealing the show entirely.

In col­lab­o­ra­tion with three stu­dents from Bentley Uni­ver­sity, the North­eastern duo built a super­com­puter within the nec­es­sary finan­cial and power con­straints laid out by the Stu­dent Cluster Com­pe­ti­tion. And then, over a 48-​​hour period last week, they let their com­puter run five data-​​intense appli­ca­tions in a neck-​​and-​​neck race to the win.

Seven other teams, hailing from as far away as Aus­tralia, also com­peted in the chal­lenge. But one thing set Northesatern’s approach above all the rest. “We chose the most uncon­ven­tional hard­ware out of all the other teams,” said Shah. Instead of building their com­puter with cen­tral pro­cessing units alone, they also used graph­ical pro­cessing units.

Both Shah and Swamy had four years of under­grad­uate research expe­ri­ence in Kaeli’s lab where they learned the ins and outs of the GPU. This spe­cial­ized elec­tronic cir­cuit was designed for pro­cessing images, as its name implies, but in Kaeli’s lab they’re also using it to process mas­sive data sets in par­allel. The approach didn’t only give them good power effi­ciency, it also brought their costs down.

There used to be some­thing called the Top 500 super­com­puters,” said Swamy. “But they were extremely power hungry. So now there’s the Green 500, too.” In keeping with that over­ar­ching shift in the com­mu­nity toward more effi­cient high per­for­mance com­puting, this year’s com­pe­ti­tion had two tracks for the first time ever: one restricted only by power, and one restricted by power and cost. Shah and Swamy’s team entered the latter, or “com­modity” track.

To achieve high per­for­mance without spending a lot, they used a low-​​cost advanced plat­form called an APU, or accel­er­a­tion pro­cessing unit. Made by Advanced Micro Devices, the APU com­bines both CPUs and GPUs on a single board. The approach won them first place in the com­modity class  but it also stood up rather well against the stan­dard teams: When com­peting against com­puters that cost nearly a mil­lion dol­lars, they said, the Northeastern/​Bentley com­puter came in fourth place.

Team open compute hard at work. Photo via David Yates.

Team open com­pute hard at work. Photo via David Yates.

While the actual com­pe­ti­tion was a har­rowing 48 hours of sleep­less­ness, spiked with moments of hair-​​pulling stress, like when they dis­cov­ered in the middle of one run that it wouldn’t finish in time, the whole endeavor involved a much larger commitment. The team spent months working on their com­puter, com­mu­ni­cating with the Waltham-​​based Bentley stu­dents using things like Google hang­outs along the way.

Before they arrived at the event in Denver, they’d made sure the com­puter was able to suc­cess­fully process all of the appli­ca­tions they knew would be in the com­pe­ti­tion, as well as readying it for a mys­tery pro­gram that wouldn’t be revealed until they set foot in the con­fer­ence hall.

They also made sure to keep a little bit of the computer’s pro­cessing power avail­able during every run to use for queueing up the next run. Some­thing in this com­bi­na­tion of good time-​​management skills, cross-​​institutional col­lab­o­ra­tion, and thinking out­side the hard­ware box bought them their suc­cess, said Swamy.

And while they may not have had pre­vious expe­ri­ence building com­puters from scratch, Shah and Swamy don’t plan to stop now. They found out during the com­pe­ti­tion that they’re pro­posal for the inter­na­tional super­com­puting cluster chal­lenge in June was accepted. Next stop, Germany.