by Brian Mattmiller, Wisconsin Week, November 16, 1994, p. 12
MADISON - A new generation of supercomputers capable of solving staggeringly complex problems -- from mapping the human genome to designing new drugs -- is being advanced by a team of University of Wisconsin-Madison scientists.
A $2.4 million federal grant awarded this fall to a UW-Madison computer science group places the university at the forefront of research into parallel supercomputing. The three-year grant from the Department of Defense's Advanced Research Projects Agency (ARPA) will help the UW-Madison team develop new computers that are faster and more adept at organizing vast volumes of data.
The grant brings UW-Madison's total federal support in this area to $4.5 million.
Parallel supercomputers operate by using many processors to work simultaneously on a single complex problem, where a conventional supercomputer uses only a few. The processor is a computer's mathematical engine, and the new parallel machines can have literally thousands of them working in tandem on specialized tasks.
The end result: Parallel computers can solve problems once too large for conventional computers.
"In parallel computing, speed is the only thing that matters," said Barton Miller, a computer science professor and one of four investigators in the project. "Your customers are people who have really big problems, where millions of computations per second is not enough. You're trying to reach billions."
Researchers Miller, Mark Hill, James Larus and David Wood, all computer science faculty, have two different projects in the works, nicknamed "Paradyn" "Wisconsin Wind Tunnel". Paradyn focuses primarily on increasing computer speed by developing tools that will automatically isolate the slowest parts of a program, and give a programmer precise information of the cause of the slow-down.
Wisconsin Wind Tunnel, a name which alludes to the wind tunnels used in aeronautics simulation, aims at developing a model to simulate new ideas in parallel computer software and hardware without the expense of building a prototype.
"We're trying to change the computers that parallel computing vendors will be selling in five years," said Hill. "We can make them faster, make the processors interact more effectively and make the software more reliable."
Parallel computers are already doing remarkable things in laboratories and industry. Miller said American Express has a parallel computer that offers lightning-fast profiles of customer spending patterns and serves as an early-warning system for credit card theft.
The computer can analyze receipt patterns, and notify the company of any radical changes in a customer's spending. That's often a red flag for credit card theft. With a follow-up call to clients, the company has been able to alert customers before they realize their card's been stolen, he said.
American Express is using the same model of computer UW-Madison owns -- a Thinking Machines model CM-5, which when purchased in 1991 was one of the fastest in the world.
The research applications are equally sprawling. Parallel computing could provide a timely breakthrough in sequencing the human genome, which is comprised of some 3 billion distinct chemical bases. It also can be used to greatly increase local precision in global weather forecasting.
Miller uses an employee metaphor to describe the drawback of parallel computing -- getting all the processors to operate in sync. If an office of 100 employees must meet continually to keep everyone fully informed on a project, the work will take that much longer to complete.
"The inefficiency of the organization grows along with the number of people who need to communicate," Miller said.
Likewise, the more processors one adds to parallel computers, the less efficient they become. Miller said the research is focusing on ways to make each processor function more independently of the others, while still sharing memory.
Their solution is called "fine-grain distributed shared memory," which, plainly put, allows each part of the computer to quickly share data from other processors. This advance creates the illusion that each processor contains memory for the entire computer.
Although cost puts the technology out of reach for most businesses and research efforts, the UW-Madison group is also looking at developing more affordable variations on the theme. A technology called "Cluster of Workstations" (COW) would allow companies with work stations or networked personal computers to run parallel software, without the costs of a new system.
"Parallel computing will not succeed with just the existence of these big machines," Hill said. "Automobiles wouldn't have succeeded if all we made were Ferraris. We have to provide the Chevrolets as well."