For Release:  November 13, 2000

Contacts:
Mark Gordon, Chemistry Department, (515) 294-0452
Saren Johnston, IPRT Public Affairs, (515) 294-3474

IOWA STATE UNIVERSITY GETS NSF GRANT TO ADVANCE CLUSTER COMPUTING

AMES, Iowa – The National Science Foundation has awarded a Major Research Instrument grant to Iowa State University’s Center for Physical and Computational Mathematics. The MRI grant is for $300,000, including $190,000 from NSF and more than $100,000 in matching funds from the ISU departments of chemistry, physics and mathematics. CPCM researchers will use the MRI funds to improve communication technology in cluster computers – personal computer or workstation networks that can operate at speeds comparable to today’s commercial parallel computers, but for a fraction of the cost.

"Users are beginning to see clusters as a good thing," said Mark Gordon, an ISU distinguished professor of chemistry and a principal investigator in the MRI effort. "A university department or a research group won’t go out and buy a supercomputer – they can’t afford it. But they can afford to put a good cluster together. So the issue now is how do we help those people maximize cluster efficiency?"

Taking advantage of the expertise and facilities available at the Scalable Computing Lab of the U.S. Department of Energy’s Ames Laboratory, CPCM researchers will investigate interconnect solutions for cluster computers. Enhancing communication, or message-passing, between computers in a cluster – making it faster and more efficient – is the primary goal of the MRI-supported research effort.

"The combination of increasingly reliable, sophisticated computational hardware and applications software for chemistry, physics, engineering and biological sciences has really taken off in the last 10 years," said Gordon. He noted that solutions to grand-challenge problems of the next decade, such as the design of new materials and catalysts, the development of viable methods for environmental remediation, and the search for the origin of life, will depend on state-of-the-art computational hardware and applications software to take advantage of modern computers. "The high-performance computing environment of the future will undoubtedly include scalable cluster computing as a component," Gordon said. (Scalability refers to the ability to increase, or "scale up," computer processing power to run the same job in less time.)

"It’s exciting to have scalable hardware where you can get several computers working together to solve a problem," Gordon continued. "In the case of clusters, we want to figure out how to best manage the hardware – how to get computers talking to each other with a minimum amount of overhead – that is, a minimum amount of time used in communicating and a maximum amount of time in actually doing the calculations. That’s what the MRI grant is all about."

The enhanced communication Gordon is talking about is all governed by the switch that sends information from one computer to another in a cluster. Currently, the Scalable Computing Lab has a $65,000 Gigabit Ethernet switch on loan from Packet Engines that directs messages to a cluster of 22 IBM dual-processor Power 3 systems. (Each system, or node, has two central processing units for a total of 44 CPUs.) IBM awarded 15 of those systems to ISU in a 1999 Shared University Research grant. With the aid of Department of Energy/Ames Laboratory funds, the cluster was expanded to 22 dual-processor systems. It is this IBM cluster and the Gigabit Ethernet switch governing it that CPCM researchers will work with in the MRI grant effort to optimize internode communications.

"The Gigabit Ethernet switch is supposed to be able to send one gigabyte, or a thousand megabits, of information per second from node to node in a cluster," said Gordon. "But it only works at about 30 percent efficiency." He noted that Brett Bode, an SCL researcher and a co-principal investigator on the MRI grant, discovered that as the number of bytes of information being sent per second is increased, efficiency goes up from 30 percent to 80 percent. "That’s the good news," said Gordon. "The bad news is that few companies, if any, are making switches that allow you to send sufficiently large amounts of information. Switch technology is really in its infancy. Companies are not designing switches that are made to do real science."

In an effort to remedy the "bad news," CPCM researchers involved in the MRI grant work are already investigating ways to optimize communication throughput with the Gigabit Ethernet switch as well as with alternative switch technologies.

"CPU power doubles about every 18 months, while communication speed doubles only every three years," said Gordon. "So it’s clear that the true bottleneck in advancing cluster computing lies in the communications. The MRI grant will help us find ways to improve interconnect technology and utilization."

The Center for Physical and Computational Mathematics is a member of the Institute for Physical Research and Technology, a network of research and technology-transfer centers and industrial-outreach programs at Iowa State University.

###

Return to Index

Last updated: 2/07/2001
mjg

Home | Disclaimer