Dumpster, dumpster spare that junk -- it might be a supercomputer

By Bill Steele

Don't throw away that old computer. Cornell students Bryan Kressler and Nick Burlett might be able to use it to make a supercomputer.

Junior Bryan Kressler, left, and sophomore Nick Burlett met in juggling club and went on to become partners in building a cluster computer -- behind them -- which consists of several old, slow computers mounted in an unimpressive rack. Just as jugglers keep many balls in the air at one time, a cluster computer runs several calculations simultaneously. And for computing jobs that can be broken into smaller pieces -- like the graphics program running on the monitor beside Kressler -- the result is high-speed computing at very low cost. Nicola Kountoupes/University Photography

Using a few old computers donated by the Mitre Corp., along with castoffs from Cornell loading docks, the two students have assembled a "cluster computer" in which the whole is much faster than the sum of its parts. It's not a supercomputer yet, they say, but expanding it is next year's project.

"It's really quite amazing, and the Mitre folks are extremely impressed," said John Belina, lecturer in electrical and computer engineering and assistant director of Cornell's School of Electrical and Computer Engineering. Belina advises Kressler, an electrical engineering junior, and Burlett, a computer science sophomore, on the project. A "cluster" is a popular new approach to supercomputing. The central processing units of conventional computers are always being made faster, but the laws of physics eventually will place limits on how fast you can push electrons through wires. So engineers have turned to parallel processing, where many calculations run at the same time on separate processors.

The Cornell Theory Center (CTC) operates clusters with up to 256 processors, and the cluster operating system Cornell developed is being used by several companies. But a positive side effect of cluster computing is that, since individual processors don't have to be all that fast, they don't have to be all that expensive. Cluster computing is making supercomputing affordable enough to make it practical as a student project

"We were looking for a project, and John Belina wanted a cluster," Burlett explained. "He has a group of students working on a program that analyzes the human electrocardiogram. It currently runs really slowly on Matlab." (Matlab is a popular mathematical analysis tool.) Both students had been interested in cluster computing, but, "it's not something that's generally accessible to undergraduates," Kressler said.

Mitre donated six Intel Pentium II machines with speeds ranging from 200 to 400 megahertz (MHz). With those machines, a couple that had been donated previously and the aforementioned loading dock requisitions, they now have a cluster computer with eight nodes, plus one other computer acting as a sort of traffic director.

While the CTC clusters use the Windows operating system -- with licenses donated by Microsoft -- Kressler and Burlett use FreeBSD, a freeware version of Unix. "We went with free software because we don't have any money," Burlett explained simply.

The idea of building low-cost clusters is certainly not unique to Cornell, he added. The ultimate example, although not a student project, is the Stone Soupercomputer at Oak Ridge National Laboratory. Just as stone soup is made up of contributed food, the Stone Soupercomputer consists entirely of castoff lab computers. "They were doing climate analysis and had no budget for computers," Burlett explained. "It's not low-cost computing, but no-cost."

"Right now I don't think you could call what we have a supercomputer, but it has the potential to be developed into one," Kressler said. With programs written to take advantage of parallel processing, Burlett said, their machine currently is about as fast as a 2000 MHz Pentium.

"It's making the processors all work as a team that's the challenge," Belina said.

They have dubbed their creation "Deep Red," reflecting Cornell's "Big Red" nickname and taking off on IBM's "Deep Blue" cluster.

Early test runs have used a program Kressler and Burlett wrote themselves that generates fractal images. It's ideal for parallel processing, Burlett said, because each pixel on the screen can be computed separately. The fractal program takes about five minutes to generate a high-resolution image when running on a single node, and about 30 seconds on the cluster. The students are working on an interface that will let the cluster run Matlab, which, Burlett said, will give them "pretty pictures," as well as helping their adviser's electrocardiogram project.