Building a Supercomputer In a Flash
San Francisco effort falls short of speed goal, but ignites imaginations
A flash mob is a group of strangers, organized over the Internet, that comes together briefly, acts out a predetermined scenario--often something as simple as clapping for 30 seconds--and then disperses.
A supercomputer is a very fast and powerful computer that outperforms most mainframes, at a cost, typically, of millions of dollars.
Can a flash mob build a supercomputer?
It can sure try, and so it did on 3 April at the University of San Francisco's Koret athletic center. The goal was to run the Linpack benchmark, a standard method of assessing the speed of supercomputers, to achieve a speed of at least 403 gigaflops (billion floating-point operations per second). This would be fast enough to earn the system a place on the list of the world's fastest 500 supercomputers.
The goal was not reached.
"That's why they call it research," Patrick Miller, one of the organizers of the effort, told IEEE Spectrum. The group reached just 180 gigaflops before it was time to pack up the computers and go home.
But would he call it a failure? No way.
"We built a supercomputer, tested it, and tore it down in six hours! How can I be disappointed?" says Miller, who is a part-time lecturer at the University of San Francisco (USF) and a full-time researcher at Lawrence Livermore National Laboratory [see photo, " Waterfalls of Cable"].
The idea first popped up in a USF graduate class taught by Miller. The students were discussing how they might build a supercomputer that could make the list of the world's top 500 supercomputers. Then one student suggested creating a flash mob event by posting a notice on Craig's List, an Internet bulletin board.
The USF group posted the announcement and then began scrambling to put together the high-speed network cables, switches, and software that would be needed.
In five short weeks, USF's flash mob supercomputer, quickly dubbed FlashMob 1, evolved into a major event. Hundreds of people preregistered, many promising to bring desktop computers, laptops, or whatever was at hand [see photo, " Computer Babble"]. Corporate sponsors jumped on board. Network switches were borrowed from Foundry Networks, in San Jose, Calif. Bits and pieces of open-source software were downloaded and cobbled together.
On Wednesday, just three days before the Saturday event, the software set was completed and burned to 1400 CDs. On Friday, students hooked up the first phalanx of computers--about 150--from the university's computer lab.
At 8 a.m. Saturday, the geeks began to arrive, carrying their computers into USF's basketball arena. Each was handed a disk and a checklist, then directed to a spot amid rows of tables, down which rivers of cables poured like waterfalls from the networking switches.
"I heard about it on Slashdot," said Alex Snell from Mill Valley, Calif., as he wrestled his tower PC onto a table. "I'm a computer freak, and I wanted to socialize with other computer freaks. And I've always been interested in the big end of computing."
"For us geeks, this is fantastic," said Igor Ranitoviz, a software engineer with Internet Archive in San Francisco.
It wasn't all young geeks. Etienne Handman, the chief technology officer of E-loan in Pleasanton, Calif., cleaned out the company's stash of spare computers and loaded 100 PCs into a van. "This event is going to prove that ordinary people with ordinary computers can build a world-class supercomputer," he told Spectrum.
There was a sense of history about to be made, and this attracted pioneers in big computing who didn't participate but simply wanted to be there, including Gordon Bell of Digital Equipment Corp. fame and Gene Amdahl, designer of the first IBM Corp. mainframe.
Sometime after 11 a.m., with 669 computers hooked up, the gym was cleared of people, and the benchmarks began to run. It was a dicey proposition all along, because the standard supercomputer benchmarks are not fault tolerant; if one processor goes down, the program fails.
And processors did go down, or simply couldn't keep up with their peers. Typically, these were older computers that barely met the requirements. After they were pulled off the network, the program was launched again, and again. Eventually, only 256 computers were left in the final run, and the speed was less than half the amount that had been hoped for.
"We had enough machines on the floor to get the 403 gigaflops that I wanted, but some of the machines were broken," Miller said. One problem was intermittent failures of the network cards in the various computers.
"But the thing basically worked," he said. "One hundred eighty gigaflops is a pretty big number--we built a supercomputer. And we have four groups ready to plunge immediately onward, including one high school, a group of small Puerto Rican universities, a team from the American Chemical Society, and a university in Australia."
USF is making the software set available as a free download to anyone who wants it. For more information, see http://www.flashmobcomputing.org.