Cluster of Pentium processors makes a supercomputer at a bargain price

Kim Alday, of Wang Global, the international computer networker, helps install rack-mounted Dell Poweredge 6350 servers that make up the AC3 Velocity Cluster, in Rhodes Hall in early August. Nicola Kountoupes/University Photography

By Bill Steele

You may have a piece of a supercomputer on your desk -- that is, if you can get together with a few friends.

A project at Cornell has linked a cluster of 256 Intel Pentium III microprocessors together to act as a supercomputer, one of the largest "tightly-coupled" systems of its kind so far, using the largest hardware switch ever made and new control software written at Cornell. Most important, it may be the most cost-effective supercomputer around. Since the system is built entirely with off-the-shelf components, such a cluster could easily be built almost anywhere and used for many scientific and business applications, Cornell experts say.

"Just as your PC is getting cheaper, this is going to drive the price of supercomputing down," said David Lifka, associate director of the Cornell Theory Center (CTC), which makes supercomputing facilities available to the university's scientists. "The thing that makes this machine special is that we used all commodity-based (off-the-shelf) components," Lifka said. "We can show that this machine is easy to replicate, for commercial, computer science and computational science applications."

The Cornell installation doesn't consist of a room full of desktop PCs, although in theory, Lifka said, the technology could even be used to link all the desktop computers in an office together for special number-crunching jobs. The machine, called the AC3 Velocity Cluster, is made up of 64 rack-mounted Dell Poweredge 6350 servers, each incorporating four Pentium III chips and running the Windows NT operating system. The servers are mounted in racks of eight and communicate with one another at 100 megabytes per second through a cLAN Cluster Switch made by Giganet Corp. of Concord, Mass.

Jobs running on the system are managed by software called the Cluster ConNTroller, written at Cornell over the past two years. The software has been licensed for commercial use to MPI Software Technology Inc. of Starkville, Miss.

The resulting system is capable of a speed of about 122 gigaflops -- 122 billion floating point calculations per second. The supercomputer that has been the centerpiece of CTC up to now, a parallel-processing IBM SP, runs at about 76 gigaflops, Lifka said.

The cluster system was assembled at a cost about one-fourth to one-fifth that of a conventional supercomputer, Lifka said.

At present the cluster system is in its early user testing phase, with about 20 Cornell research groups porting applications they had previously run on the IBM supercomputer, including a hugely complex simulation of protein folding.

The cluster project is the result of a collaboration with Dell, Intel, Microsoft and Giganet, who are all members of CTC's Advanced Cluster Computing Consortium (AC3 ), established to bring together businesses, higher-education institutions and government agencies interested in the further development of cluster computing. AC3 also includes 15 members from the software industry. Consortium members will receive technology briefings, training, and consulting services from Cornell's Cluster Computing Solutions Group, which can provide assistance in planning commercial off-the-shelf systems.