PCs Hit the 'Outer Limits'

Application Limits

But distributed computing isn't for every job, says John M. Old, director of information management for worldwide exploration and production at Texaco in Houston.

"The SETI project lends itself to breaking the data into small, independent chunks, which makes the parallel computing fairly simple," Old explains. Unfortunately, not all data can be segmented that way, and many projects require complex communication among processors.

McNett acknowledges that there are plenty of things an IBM RS/6000 can do that a distributed network can't. For example, weather prediction is difficult because the data is very interrelated. Distributed computing is better at jobs such as animation rendering, in which each of the 30 frames per second that go into a movie like Toy Story are separate tasks that can be distributed among thousands of computers.

With those kinds of jobs in mind, the folks at Distributed.net are considering a commercial spin-off. At present, Distributed.net's machines are equivalent to 42 144-node RS/6000s, the fastest computers on the market, at a net cost of about $120 million (based on the floating-point speed of the RS/6000 and the Pentium II/266 PC, the average computer on the distributed network).

If the SETI project rallied 2 million computers by word of mouth, imagine what a company that was willing to pay for your PC's time might accomplish. That's exactly what Jim Albea, chief operating officer at ProcessTree Network, in Madison, Alabama, was thinking in January when he set up a Web site soliciting computers for the April launch of what he claims is the first commercial venture in the field.

Despite the potential, there are problems that have to be solved before massively parallel Internet computing can work commercially, McNett says.

The biggest hurdle is security. An oil exploration company considering the mineral rights to some land might gain a lot of efficiency by divvying up the analysis of the geologic data across the Internet. But what's to stop a competitor from setting up machines in the network and gleaning some insights from the data? And what about would-be saboteurs in the network, bent on ruining a project for competitive or malicious reasons?

Another concern is that if people can modify the software's behavior, they can affect the project's integrity. Finally, McNett says, massively distributed computing calls for a business model that has yet to gel.

"Are you going to send 18-cent checks to 100,000 people every month?" he asks.

Albea says he thinks ProcessTree has solved most of the technical and business problems. Even though ProcessTree hasn't yet set a pricing plan, CEO Steve Porter offers a ballpark figure of about $1,000 for the equivalent of a year's worth of CPU power from a Pentium II/400.

The company may pay in the range of $10-to-$20 per month per computer--and even more for large-volume volunteers such as businesses. Payment will likely be in credits with an online retailer or service. For example, a participant might get discounts on Internet service in exchange for running the software.

Since its site debuted in January--with virtually no advertising--ProcessTree has lined up more than 35,000 users representing more than 70,000 machines.

"We are the largest body of available commercial computing power in the world right now," Porter says. "You can't get anything that can go faster than we can, and we get faster every day."