November 27, 2000
by Bruce Upbin
Nancy Davis didn't want to spring for another $10 million supercomputer, but she has needs. As a technology manager at Pratt & Whitney, a manufacturer of jet engines, she supports engineers who must simulate-in 3-D-the structural and fluid effects of superhot gases rushing at 685 kilometers an hour through engine turbine blades spinning at 9,000 revolutions per minute.
One night a few years ago, after the staff had left, she surveyed the company's design center in East Hartford, Connecticut, and counted hundreds of powerful Unix workstations sitting ready but idle. What wasted resources. So her staff rewrote software to chop structural analysis into small chunks and distribute them to 5,000 workstations to process overnight.
The result: a virtual supercomputer that's capable of crunching 6 trillion floating-point operations a second-six teraflops, equivalent to one of the ten most powerful computers in the world. "It's essentially free, since we already use them during the day," Davis says. It works so well that she plans to add 13,000 PCs to the effort.
Lashing together multiple processors to create a virtual supercomputer is a popular parlor trick in academia and government research. Now such "distributed processing" is beginning to solve thorny business problems, from complex graphics through risk analysis to design.
J.P. Morgan, Intel, Microsoft, Boeing and others harness their far-flung PCs to perform-on the cheap-intense computation that was previously either impossible or too costly. Now a slew of new companies are racing to enlist thousands of individual PC owners to join their "grids," aiming to rent out this parallel power to computation-hungry clients for a fee.
Such schemes gained fame with the project known as SETIAtHome (SETI stands for Search for Extraterrestrial Intelligence). Started in late 1998, the space-alien hunt has recruited the owners of 2.5 million PCs, tapping their machines' idle processing time to search small chunks of radio telescope data for patterns that might signal intelligent life elsewhere in the cosmos. Despite 25 teraflops of computing power, SETIAtHome has not yet spotted any little green people. But it has inspired other schemes, including a Stanford University study of how protein chains fold in on themselves and a global weather- forecasting model at the Rutherford Appleton Lab in Britain.
Cool stuff-but now serious money is pouring into the concept. Kleiner Perkins, the venture-capital firm, invested $6 million in Centrata, which is in Menlo Park, California. Entropia, in San Diego, California, raised some $7.5 million from Mission Ventures and others in May. United Devices, in Austin, Texas, raised $13 million from Softbank and Oak Investment Partners in August.
The new ventures typically sell software that lets a company lash together hundreds or thousands of PCs in-house; they also hope to line up home PCs for corporate tasks, although the outlook at that end of the market is questionable.
The parallel push underscores two truths in computing: first, that today's Pentium III computers are so embarrassingly overpowered that users tap only 5% of the real muscle available; second, that sharing "peer-to-peer" functions is rapidly becoming commonplace, from Napster music downloaded by 38 million people to content, bandwidth and storage.
The cost savings can be enormous. Intel figures that it saved $500 million in hardware costs over ten years by using a homegrown application called NetBatch, which sits on thousands of workstations around the world. When a designer is ready to test part of a new chip layout, his workstation uses NetBatch to find all the other machines on the network that are idle, whether they are in Ireland, China or the U.S., and feed the job to them. Intel's workstations are now used 80% of the time, up from 30%.
"This was one of our little secrets we didn't want to disclose," says Robert Knighten, the peer-to-peer evangelist with Intel's Enterprise Architecture Lab, which develops software for big systems. "But in the past few years a lot of companies have developed similar applications."
For the past year J.P. Morgan has been simulating trades in interest rate derivatives on 250 PCs in its London offices. It had always used powerful Sun workstations for such risk modeling, but that was getting too expensive. With the cheap new muscle that's been added, J.P. Morgan can do complex calculations that were prohibitively expensive and time-consuming, such as running a trade through scenarios for both interest-rate fluctuation and market volatility, rather than looking at just one at a time.
The investment bank had looked at renting processing cycles from supercomputing centers in Illinois or Minneapolis, Minnesota, but it chose to keep the work in-house. "We would be seriously interested in doing it over the Net-with good encryption," says Michael Liberman, a J.P. Morgan vice president. "Our problem is limitless; there will always be more risks we want to analyze."
In September BEA Systems, the big e-commerce softwaremaker in San Jose, California, started using linking software from a shop called Popular Power to test the new version of its Weblogic Server software. BEA ordinarily buys or rents server time to simulate shoppers bombarding a web store running Weblogic.
Web hosts can charge up to $200,000 to run one of these tests. So instead BEA installed Popular Power's software and a testing program on 50 powerful Windows NT boxes in its San Francisco offices. Each can simulate 50 to 100 shopping transactions at once, whenever a PC is unused for six minutes or so during the day and continuously at night. "Right now we're just proving the concept, but the savings are in the tens of thousands of dollars," says Boris Chen, a BEA manager.
Boeing uses an even more complex distributing program, called Legion, from Applied MetaComputing. It hooks up hundreds of computers around the world to get them talking to one another at the same time to calculate acoustic and electromagnetic effects on the fuselages of fighter planes when they are in steep angles of attack. Unlike a SETIAtHome kind of project, in which data are crunched independently on separate PCs, Legion can run tasks that require some synchronization ("Don't do that until I do this.").
The idea of new businesses seeking to tap the slumbering processing power on thousands of home PCs sounds enticing, but the models are unproven. Although the at-home pitch worked well for the SETI project, businesses are likely to keep their projects behind company firewalls.
Handing off sensitive corporate data to unknown PCs over the public Internet poses serious technical and security issues. PCs vary in power, operating systems and availability. If a participant shuts down his PC during a crucial bit of computation to go on vacation for two weeks, the work it was doing must be reassigned. Limited bandwidth can jam up traffic, forcing some slices of processing to wait in long queues. And while data can be encrypted in transit, they have to be decrypted while a hard drive does its work. That invites prying eyes to piece operations together.
On the demand side, the bulk of customers currently tapping into these public grids are nonprofits in search of freebies: AIDS researchers at the Scripps Research Institute, modeling the molecular interaction of new drugs and the body's cells; chemists studying protein-folding at Stanford University.
Amassing a huge supply of clock cycles won't be easy, either. SETIAtHome had no problem recruiting: The spacey project became a competitive sport for engineers. So the companies pushing home-PC projects resort to bribery.
DataSynapse issues "flooz," a form of digital currency. Porivo Technologies offers a shot at a free mountain bike. Popular Power and other entrants mull a microtransaction system that will accumulate monthly payments for those who sign up.
For PC owners, this won't be a get-rich-quick scheme; by some estimates they will be lucky to earn $10 to $15 a month. Adam Pavlacka, a vice president at Mithral Communications & Design, which develops power-sharing software for businesses, cautions: "The [U.S. Internal Revenue Service] is going to laugh if you do this to deduct the cost of your computer."
Maybe so, but this new power-sharing offers companies a rare opportunity in high tech-a chance to tap sleeping PCs that were too powerful to begin with, and to make the most of what languishes on employees' desks.
© Mithral Inc.
All Rights Reserved.
Mithral® and Cosm® are trademarks of Mithral Inc.