A team of computer scientists linked hundreds of individual PCs into a temporary supercomputer on Saturday. It take the potential of supercomputing out of the labs and puts it in the hands of independent researchers.
But while the organizers of the FlashMob Computing project succeeded in one respect -- building the world's first ad hoc supercomputer out of a gym full of loaner Wintel machines -- they failed to crack the world's top 500 list of largest supercomputers.
"The bottom line is, the (computing) rates that we got were pretty darned impressive, and we're pretty happy with what we got in the time we had," said Greg Benson, an assistant professor of computer science at the University of San Francisco and one of the primary organizers of the FlashMob event. "Everybody involved was just really happy in the end."
Saturday's event, at USF's Koret Gym, was a geek super-fantasy come to life. All morning, PC owners streamed into the building, dropping off their machines, which were then plugged into a matrix of hundreds of others.
In the end, the room resembled a government auction, or perhaps a gargantuan Homebrew Computer Club meeting: 700 PCs of all different makes, models, shapes and sizes; laptops and desktops and even machines with little more than a microprocessor and no box, all lined up, row after row.
"When you're begging and borrowing, you have to get what you get," said Pat Miller, another leader of the project and a lecturer in computer science at USF. "It's kind of a stone soup."
And indeed, those who brought machines for inclusion in the project were excited at the prospect of taking part in something that could forever alter the face of large-scale computing.
"I think it's really cool," said Daniel Baker, who toted three computers to Koret Gym Saturday. "It's an opportunity to be a participant in science and do everything I can to perpetuate computing.... If supercomputing can be deployed on much greater and cost-effective levels, how much more available they can become to solving important problems."
FlashMob Supercomputing is USF computer science student John Witchel's master's thesis project. He says proving that it's possible to utilize a large number of computers at little or no cost allows groups of people who wouldn't normally have access to supercomputing power to address pressing computational problems, such as studying global warming or breast cancer.
"We're trying to show that we've got a usable form of (supercomputer)," said Witchel, who wants to "get it out of the laboratories and into the hands of the people."
Witchel and his colleagues were inspired, in part, by the recent construction by scientists at Virginia Tech of a supercomputer using 1,100 Apple G5s. The USF team imagines smaller supercomputers could be constructed by any group with a couple dozen or more Wintel machines, using the software they wrote for the event.
"The temporary nature of the FlashMob ad hoc supercomputing, is that you can put it together whenever you need one," said Miller.
Event organizers predicted they would get around 1,000 loaners for the project. They hoped with those machines -- all linked together through several hardwired hubs -- they would be able to create a single computing entity big enough to make the list of the world's 500 biggest supercomputers.
To do so, they would have had to get a sustainable performance of at least 403 gigaflops. (A gigaflop is a billion floating-point operations per second.) The FlashMob team achieved a momentary peak performance of 180 gigaflops and a final result of 77 gigaflops, said Benson.
From the beginning Saturday, the project was beset by problems. First, fewer people showed up with their PCs in hand than were expected. That was evident when an entire section of the gym set up to link PCs remained empty. Then, problems with failing computers caused a late benchmark test to come to a halt at about three-quarters completion.
In the end, the one completed benchmark utilized only 150 of the 700 PCs in the room, said Benson.
"Since this was the first time, we didn't have a lot of time to figure out what we had," said Benson. "I'd say that was the biggest challenge looking back now -- figuring out how to make the resources we had work together."
Throughout the morning, a large team of volunteers worked to get the computers up and running. In one corner of the room several of them ran a makeshift "hospital" to fix problem machines. At one point, Francisco McGee, a volunteer, hustled a laptop from the hospital and plugged it into the matrix.
"Yes!" he shouted, as he saw it come to life. "Got that [censored] online." Asked what had happened, McGee said volunteers had given the machine "CPU CPR."
But as morning turned into afternoon, it became clear that the organizers were not getting the performance they had hoped for. Initial optimism at the prospect of making the Top 500 list turned quickly into hopes they would at least end up in the ballpark.
Miller had originally hoped the team could achieve 550 gigaflops, given the belief they would have 1,000 machines on hand. But, "that's going to be a push," he said around midday. "I'm hoping for 400. I think we've got 300 cold."
However, because of an inability to keep a single failing computer from bringing the test to a halt, the team, pressed for time, resorted to using a much smaller number of machines, and even Miller's safe 300 gigaflop expectation ended up unrealistic.
For Miller, Benson, Witchel and the others involved, the real point of Saturday's exercise was to get a lot of people involved in making history. And that they did.
"It's exactly how I envisioned it," said Witchel of the mood in the gym Saturday, "just tons and tons of excited people who want to make a difference."
Some in the supercomputing community felt the FlashMob project was a nice beginning, but that people shouldn't expect to be running out to set up their own supercomputers right away.
"It's maybe a long way to go to accomplish the vision the organizers have of an easily accessible supercomputer that could be used by everybody today," said Horst Simon, the associate laboratory director for computing science at Lawrence Berkeley National Labs and a member of the Top 500 project. "But it's the first step, and somebody's got to start it, and that's why I think it's good."
For his part, Benson believes he and his team will take the lessons they learned Saturday and use them to improve upon their performance in the future.
"My vision is that we're going to work on the software," he said, "now that we have a little less pressure. The software we developed, we did in five weeks. Now we're going to stand back and say, 'How can we make this better?'"
You can see the original article here...