College knits off-the-shelf Macs into third-fastest computer
Sunday, November 16, 2003
BLACKSBURG, Va. -- It comes from the bargain basement of supercomputers, but the performance is nothing short of astounding.
At Virginia Tech, 1,100 Macintosh PCs are stacked like library books on metal racks that students helped arrange in return for football tickets and pizza.
The cluster of off-the-shelf G5 Power Macs was assembled in a few weeks for about $7 million. The custom supercomputers that research labs use for weather and weapons simulations and other highly complex projects can cost more than $200 million.
But Virginia Tech's "Big Mac" -- as students have begun calling it -- is about to rank as the world's third-fastest supercomputer, at 10.3 trillion operations per second.
"It's really quite impressive," said Jack Dongarra, a computer science professor at the University of Tennessee who compiles an annual list of the top 500 supercomputers. "They're now competing in terms of performance with our foremost research facilities."
Dongarra, who will release the 2003 rankings Monday, said he expects the only faster supercomputers to be one in Japan and a computer at Los Alamos National Laboratory.
But some experts believe the supercomputer business need not fear home-brew rivals.
"These clusters have their trouble points," said Ed Seidel, who supervises a computer cluster at Louisiana State University. "You know how often your own PC fails sometimes? Just think of 1,000 computers. Custom-built systems are more reliable at times -- they were engineered to be supercomputers."
Even so, Virginia Tech's computer cluster dwarfs the power of other systems that have been created by engineers who linked together hundreds of commercially available processors.
Each of the 1,100 Macs has two IBM PowerPC 970 microprocessors that are based on a relatively new 64-bit design, which means they process data in chunks of 64 bits -- a method exponentially faster than older, 32-bit technology. The processors, running Mac OS X, are connected by a high-speed network called Infiniband that allows them to break up major calculations and analyze each part at the same time.
Virginia Tech quantum chemist Daniel Crawford said Big Mac will shorten the time he spends building computer models of chemicals from years to just days.
"It opens an entire new area of chemistry," said Crawford, who will be one of the first to use the computer in December. Before Big Mac, Crawford would have to wait in line to use part of another computer. "You could spend your entire career on just a few calculations."
Project leader Srinidhi Varadarajan said a cluster like this was inevitable as personal computing equipment got more powerful. Other universities will likely follow with even faster clusters, he said, drawn as Virginia Tech was to the cheap PC components and the pressure to enhance their academic reputation despite severe budget cuts.
Dan Powers, vice president of grid computing strategy at IBM Corp., says he's not worried about such off-the-shelf PC clusters threatening sales of traditional supercomputers.
For one thing, Powers says, the Big Mac project took a lot of labor that Virginia Tech was able to get for free. Clustering computers into a supercomputer requires not just physically linking their cables and input-output systems but also tying the whole thing together with a specialized "back plane," a high-speed network that relays data throughout the system.
Powers said almost all supercomputer users would rather pay companies like IBM to do all that than try to build their own. Users of supercomputers from companies like IBM can expect to shell out $100,000 to $10 million, though highly advanced models for nuclear simulations range closer to $200 million.
Virginia Tech began seeking a supercomputer in the spring. Hassan Aref, the dean of engineering, said school officials approached Apple Computer Inc. after another computer company turned them down.
Apple, which has had little presence in the supercomputing industry, was reluctant, he said, but "we persuaded them that we could make it work."
The school spent about $5 million on the computers and another $2 million renovating the building that would house them, Aref said.
In September, Varadarajan and about 160 student volunteers spent days removing the computers from their boxes, installing communications software and loading them on rows of liquid-cooled metal racks.
For the past six weeks, Varadarajan and his staff have been running test programs on Big Mac, teaching it to think faster. The supercomputer already has added about 3 trillion operations per second to its initial performance benchmark, and Varadarajan expects it to be even faster when Big Mac is formally opened to university use in January.
"This is like tweaking a race car," he said.
On the Net: