FEATURES & COMMENTARY

San Diego, CALIF. — In mid-April a group of 20 Vanderbilt University physicists, biologists and computer technicians held a two-and-a-half day pizza party. At its end they had assembled a powerful supercomputer from off-the-shelf PC parts for a bargain-basement price.

“When they told me what they planned to do, I was skeptical,” admits Chip Cox, Vanderbilt’s Internet 2 director who helped bring together faculty members from the campus and medical center who need supercomputer-level number crunching. “But I was wrong. They did it, and it works!”

They named their new number cruncher VAMPIRE, which stands for Vanderbilt Multiple Processor Integrated Research Engine. It was constructed with a parallel architecture. That is, it uses a large number of processing units, or nodes, that work at the same time and are connected by a high-speed network.

Such a design can achieve tremendous speed and power by processing information simultaneously at many nodes.

Parallel processing requires users to break down their problems into small pieces that can be processed by individual nodes and then reassembled to provide the final result. This is easier to do with some problems than with others. Fortunately, a number of science problems lend themselves to such an approach and the tremendous cost-advantage is motivating many researchers to adapt their methods to take advantage of it.

The Vanderbilt group is not alone in such an effort. Scientists, engineers and computer technicians at a number of universities around the country are building similar homebrew systems. “The average university system currently is about 16 to 32 nodes,” says Alan Tackett, the Information Technology Services (ITS) administrator who manages the new machine. “So VAMPIRE is considerably larger than average.”

Parallel computers built from off-the-shelf hardware have been dubbed Beowulfs, after the name of one of the first computers of this type built in the early 1990s at NASA’s Goddard Space Flight Center in Greenbelt, Md. The University of New Mexico has constructed a 500-node Beowulf and recently the University of Pittsburgh received funding for a 1,000-node machine, Tackett reports.

“Of course, we’d like to grow VAMPIRE to that size,” says Paul Sheldon expansively. Sheldon, an associate professor of physics and astronomy, has spearheaded the project along with Jason Moore, an assistant professor of molecular biology and biophysics, and Will Johns, an assistant professor of physics and astronomy.

The new Vanderbilt supercomputer was declared operational Sept. 11, following a several-month period of installing and debugging the software. The last element added was a “gateway server” that controls access to the processor array and allocates the computer resources to different jobs in an equitable fashion.

For the April party, boxes containing dozens of processors, motherboards, hard drives, network cards, and memory cards were set out in a small conference room in a campus building like techie party favors.

“It was amazing to see thousands of dollars worth of memory chips in a box the size of a toaster,” says Tim Miller, a physics graduate student. “It was like being a kid in a candy shop.”

The party-goers’ task was to assemble more than 50 stripped-down personal computers. The researchers and technicians plugged in two 600 megahertz Pentium III processors and 256 megabytes of Random Access Memory or RAM and connected a 10 gigabyte hard drive and a networking card to each motherboard. Each motherboard was installed in a metal cabinet about the size, quite appropriately, of a pizza box.

The components installed in each cabinet act as a single processing unit, or node, in the finished supercomputer. The nodes are all networked to each other and to the gateway server. The remaining part of the supercomputer is a server that operates a bank of hard drives with a total capacity of 350 gigabytes that serves as a central data storage area.

It took the group a little more than two days to assemble the 54 nodes. Individuals came and went as their schedules dictated. There were about 10 people working in the room at any given time. One of their biggest problems, they agree, was a shortage of screwdrivers.

In its current configuration, VAMPIRE cost about $90,000 to build. That is less than the price of a single node for IBM’s least expensive supercomputer, the SP-1. In addition, two VAMPIRE nodes are more powerful than one SP-1 node, says Tackett.

Homebrew projects like this are possible because computer components have become standardized commodities, Sheldon points out. “That allows us to shop around for the best price on individual components with the confidence that the pieces will all work together.”

Another critical element has been the development of Linux, an operating system that is particularly suited to this kind of application. Linux is an “open source” version of Unix, a scientific operating system originally developed by AT&T that is the standard for basic research applications. Open source software is developed, distributed, and upgraded for free by independent programmers. Using Linux saved the project thousands of dollars in software licenses, according to the researchers.

In 1994, Sheldon and his colleagues in the physics and astronomy department built a smaller “workstation farm” with a $250,000 grant from the National Science Foundation to perform the computations required by their projects, such as an experiment that Sheldon is participating in at the Fermi National Accelerator Laboratory in Batavia, Ill. that studies the behavior of quarks, the tiny particles that make up the protons and neutrons that reside in the nucleus of the atom. With the prospect of new experiments that will produce even more data, the physicists were looking for a way to boost their computer power.

Last year, Moore, whose research also requires a large amount of data processing, received a campus grant to finance the purchase of a larger, faster computer. He learned about the physics department’s workstation farm, contacted Sheldon and the impetus for VAMPIRE was born.

After their initial success, the VAMPIRE team would like to grow the system into one of the largest Beowulf computer arrays in the country. They envision doubling, tripling, quadrupling the number of nodes and linking them with new, high-speed networking capabilities that should significantly jump up the system’s performance.

“VAMPIRE provides campus investigators with an invaluable research tool,” says Sheldon. “It will allow them to compete successfully for research projects that involve intensive computation, as well as bringing increased funding specifically targeted to support computer facilities to campus.” Visit http://www.vampire.vanderbilt.edu for more information.

============================================================