boston.com Business your connection to The Boston Globe

Ready, set, compute

Pentagon driving a renewed race for faster machines

BURLINGTON -- Supercomputing was once primarily about speed, powered by drug researchers and weather forecasters who needed to process massive volumes of data quickly. Now the Pentagon is in the market for better number-crunching, and it's changing the rules of the game.

A new round of Pentagon funding has Sun Microsystems Inc. vying with IBM Corp. and Cray Inc. to build a computer that will be faster, more powerful, and more agile than any now in use. The program is called HPCS, for high-productivity computing systems. The winner will not only earn bragging rights for building the world's fastest computer, but it may transform commercial computing in the process.

"This is radical," said Steve Heller, director of SunLabs East, tucked away in the sprawling Sun campus off Route 128. "If we do this right, we're going to rock the industry."

The competition is being engineered by the Pentagon's ideas factory, the Defense Advanced Research Projects Agency, which has funded research into everything from stealth jets to the Internet. Long a benefactor of New England research labs, DARPA's mission is to keep the United States in the forefront of military and civilian technology. DARPA's latest idea is to nurture a new generation of supercomputers for national security and industrial uses. Toward that end, in 2002 it funded concept studies by five companies: Sun, IBM, Cray, Hewlett-Packard Co., and Silicon Graphics Inc.

Last month, it narrowed the field to three, awarding about $150 million in "second phase" contracts to Sun, IBM, and Cray. Over the next three years, these three companies will be researching technologies that may dramatically boost computing power, and developing prototype designs for new computing systems. DARPA will then fund full engineering and development, at amounts yet to be determined, for up to two companies.

DARPA wants not only speed, historically measured in gigaflops, but also computers that are durable, easy to program, and adaptable for a range of applications. The project may lead to a new "value-based metric system" for measuring computer performance, said Robert B. Graybill, program manager in DARPA's Information Processing Technology Office.

While setting the overall goals for the project, which is meant to influence the computing environment by the end of the decade, DARPA is leaving the design up to companies.

Their challenge lies in the rarefied field of high-end computing, where multiprocessor servers crunch vast quantities of data in applications as diverse as military intelligence, weather, and drug discovery. DARPA is clearly interested in defense uses, such as intelligence gathering and weapons design.

"If we look at the future national security challenges . . . our current computers are not going to be adequate," Graybill said. But there would also be commercial spinoffs, in the design of cars and planes, for example.

"Supercomputers are a very esoteric market," said Alex Roland, a Duke University history professor who has written a book on DARPA's pursuit of strategic computing. "There's a lot of prestige involved in having the world's fastest computer, but it's not clear these monster machines are worth the effort for most applications."

Still, the US high-end computing community was jolted last year when the Earth Simulator, a Japanese computer that studies climate, laid claim to being the world's fastest. It was the first time in six years that a supercomputer outside the United States led the Top 500 list of the most powerful computers compiled by German and American researchers.

The news was greeted by some in Washington as a reprise of Sputnik, the Russian satellite that was shot into orbit in 1957 and stirred anxiety that America was losing its scientific edge. There has been a flurry of government activity over the past year aimed at rejuvenating supercomputing, and agencies from NASA to the Department of Energy are monitoring the DARPA project.

"It comes from a realization that something's wrong," said Burton J. Smith, chief scientist at Cray Inc. in Seattle. Smith said some supercomputer users have concluded that current models are still too difficult to program and adapt to various applications.

"It's possible to build systems with very high performance on paper that don't perform well in practice," he said. "They're looking for execution performance, making good use of the machine."

DARPA's emphasis on "high productivity" is a recognition that more than speed is required to achieve the breakaway innovation it is seeking. Its contractors are being urged to seek improved reliability, ease of programming, and portability.

"What we'd like to do is to build more of a general-purpose system that can morph or change itself so it can run different applications efficiently," said Michael Rosenfield, director of the IBM Austin Research Laboratory in Texas.

All three finalists remain guarded on the details of their efforts and, in the tradition of high-tech R&D, have given their projects code names.

Sun's is Hero, because it "scales to machines of heroic proportions," according to its internal website. IBM's is PERCS, for productive, easy-to-use, reliable computing system. And Cray's project is called Cascade, for the mountain range east of Seattle.

Sun, which received its $49.7 million DARPA grant last month, aims to apply its expertise in programming technologies and scalable computing toward the 36-month research effort, said Heller.

Heller said Sun's programming language, Java, is an asset because it enhances productivity. Sun is also strong in scalable computing that addresses the "rendezvous problem" of letting several processors work on a single database. Sun will have 40 to 60 researchers on the project initially, mostly in Mountain View, Calif., and Burlington, but could ramp up to 100 as the project gathers momentum.

Robert Weisman can be reached at weisman@globe.com.

SEARCH GLOBE ARCHIVES
 
Globe Archives Today (free)
Yesterday (free)
Past 30 days
Last 12 months