Something is different about Room 2-105 in the Research and Technology Development Center at the Applied Physics Laboratory. There are an awful lot of computers in there, but then that's true for many APL spaces.
Come inside. It's immediately cooler. And now you see 16 desktop computers labeled BORG 1 through 16, all connected by a mass of yellow cables to a small black box. You're looking at a supercomputer in the making.
It's the brain child of George Mou of the Research and Technology Development Center. Starting a year ago with support from the RTDC and an APL Independent Research and Development grant, Mou began putting together the hardware, software and programming model that co-workers call BORG for Brokered Objects for Ragged-network Gigaflops--in plain language, a supercomputer.
Mou is well-equipped to face his challenge. Before coming to APL in 1997, he was a Brandeis University faculty member researching parallel processing and had worked for IBM, where he oversaw performance aspects of a commercial supercomputer. At APL, he continued developing his own programming system for supercomputers called Divacon for Divide-and-Conquer--exactly what was needed when he conceived of BORG.
"As with other supercomputers, we use parallel processing to get speed and power," Mou says. "But BORG's unique software, programming model and operating principles put it in a class by itself."
According to Mou, BORG's "open" architecture makes it easy for the computer to work with different systems, hardware and computer languages. He says BORG is easy to program and it's very good at "transformations"--changing part of the program so the problem can be solved faster.
BORG is similar to but different from the Network of Workstations approach to supercomputing. "In most NOW systems there are 'slave' processors that all communicate with a 'master' processor, which can create a bottleneck," Mou explains. "With BORG, each processor can talk to every other processor, so things go faster and smoother."
While the current 16-unit BORG system costs as little as two Sun Ultra2 2300 workstations, it matches the performance of eight such stations, delivering one gigaflop (a billion floating-point operations per second).
Right now Mou and Lien Duong of the RTDC's System and Information Sciences Group are implementing BORG's programming model, developing algorithms and building applications libraries that perform such tasks as polynomial evaluation, vector reduction and matrix-vector multiplication so the computer can get to work.
"This is an opportunity for me to shape a system according to my own philosophy and vision," Mou says. "It's very satisfying--a dream come true."
BORG's first job--and a test of its abilities--will be to analyze the wave propagation characteristics of radar signals as they strike the ocean's surface. "BORG will solve problems that were too big or too hard to work on before," says Denis Donohue of the RTDC's Physics Group, which has developed a computational physics approach to solving problems involving scattering.
Next, NASA's Advanced Technology Development program has committed to using BORG to conduct a virtual spacecraft simulation. Employing a mix of real components and virtual ones generated by BORG, engineers will be able to test spacecraft performance during design and construction.
Like everyone else, BORG is going to surf the Internet. Once there, Mou says, it will be able to harness computing power from other sites throughout the World Wide Web--making BORG truly more than the sum of its parts.