Mail-Order Supercomputer among World's Fastest

Los Alamos National Laboratory Jim Danneskiold, (505) 667-1640 / [email protected]

LOS ALAMOS MAIL-ORDER SUPERCOMPUTER AMONG WORLD'S FASTEST

LOS ALAMOS, N.M., June 18, 1998 -- A supercomputer built from ordinary personal computer components is among the 500 fastest computers in the world, an international survey reported today.

The Avalon computer cost just $150,000 to build, and can compute more than 20 billion mathematical operations in a second, said Michael Warren of Los Alamos National Laboratory's Theoretical Astrophysics Group.

Avalon made the 315th spot on the 11th TOP500 list released at the Supercomputer '98 conference in Mannheim, Germany. The list is the best-known ranking of supercomputer performance.

"It's now possible for a small group of motivated people to design and build their own parallel supercomputer using off-the-shelf computer parts and easily available software," Warren said. "Only a handful of companies in the world produce a computer this fast, and the least expensive costs well over a million dollars."

Avalon is built out of 68 high-end personal computers that use the Digital Equipment Corporation Alpha microprocessor, connected by 3Com network switches similar to those found in a university department or small business. Each processor in the Los Alamos supercomputer is an ordinary PC, using the same type of memory and disk drives found in a computer on an office desktop.

"Each of these processors theoretically is capable of performing over one billion operations a second, and we bought them at consumer prices," said Warren.

But hardware is only half of the equation. Software is the hardest part of getting many processors to work together on the same problem. The Los Alamos team used an open source Linux operating system and other software available on the Internet.

"The key to the success of these machines lies in their software, and the most important part of that software is the Linux operating system," Warren explained. "Linux can be obtained at no cost through the Internet, but that is minor compared to its other advantages. In my experience, the reliability and performance of Linux has no peer.

"We have stressed Linux well beyond where one would expect it to fail, and it has performed admirably. Because it was developed as open source software, we can go to the source code and fix many problems immediately," Warren continued. "If we can't fix it ourselves, we can tap the huge pool of Linux expertise on the Internet."

While some question the reliability, complexity and difficulty of installing software on a "do-it- yourself" supercomputer, Warren and his team had no problems.

"We got most of the parts for Avalon on Friday, April 10. Three days later, the machine was computing at over 10 billion operations per second." he said.

By Wednesday, which was the deadline for TOP500 list entries, Avalon had achieved 19.2 billion floating point operations per second. The computer hasn't suffered a single hardware failure or operating system crash on any of the 68 processors during the last six weeks.

Working with Warren to build Avalon were David Neal, systems administrator for Los Alamos' Center for Nonlinear Studies, and David Moulton and Aric Hagberg, both from the Mathematical Modeling and Analysis Group.

In its short life, Avalon already has performed some significant scientific computations. One of the first simulations followed the evolution of a shock wave through 60 million atoms. The simulation ran for more than 300 hours on Avalon, calculating about 10 billion floating point operations per second.

Physicist Peter Lomdahl, who won the Gordon Bell prize for significant achievement in parallel processing using the Connection Machine 5 supercomputer at Los Alamos said the Avalon system was extremely easy to use.

"We ported our molecular dynamics code over in about a day and have been able to perform state-of- the-art simulations of shock-waves in metals that ordinarily would have required the Lab's large-scale shared-memory parallel systems" Lomdahl said. "Not only does the Avalon system run slightly faster than a similarly sized commercial system, it does it at a tenth of the cost, and is much easier to use."

Warren will use the machine in his computational astrophysics research, performing simulations of galaxies. "I am interested in simulating the evolution of the universe from its very early stages up to the present day," Warren said. "We can test different ideas about the way the universe is put together by comparing the galaxies simulated inside the computer with real observations made by the latest generation of telescopes. Avalon puts the computational power we need to do those simulations inside our own building, at a price we can afford."

In its "spare time," Avalon helped crack the Certicom Elliptic Curve Cryptosystem challenge, winning a $4,000 prize that was donated to the Free Software Foundation. The Foundation led the development of many of the software tools Avalon uses.

The code-breaking calculations ran at the same time as other large simulations, but only made progress when the computer didn't have anything else to do.

Initial funds to buy and build Avalon came from the Center for Nonlinear Studies. Other funding came from the Laboratory Directed Research and Development program and the Theoretical Division. Shi-yi Chen, deputy leader of the Center for Nonlinear Studies, said "Avalon will be used for fundamental researches in nonlinear sciences for a variety of areas, including applied mathematics, material sciences, complex systems and climate modeling."

Warren has used parallel computers throughout his career, including several which have held records as world's fastest at the time. In 1996, he built his first off-the shelf computer, Loki, which last year won the Gordon Bell prize in the "price to performance" category.

"Loki proved itself as the most cost-effective way to perform large-scale scientific simulations last year, and now Avalon provides ten times that performance for only three times the price," Warren said.

Computers using off-the-shelf technology like Loki and Avalon are called "Beowulf" computers, after the project begun by Thomas Sterling at the NASA Goddard Space Flight Center.

"Avalon is a dramatic demonstration of the long-term potential of the Beowulf model for scalable, high-end computing to perform real-world applications in science and engineering at unprecedented price-to- performance ratios," Sterling said. "Since 1994 when the earliest Beowulf systems were developed atNASA, a rapidly growing community world-wide has emerged to apply the Beowulf approach to a broad range of important problems.

"Avalon represents a new generation of Beowulf systems -- breaking new ground in performance and extending their utility to new and important areas," Sterling said.

Warren thinks that Avalon's success is only the beginning.

"In the future, I imagine hundreds or thousands of machines of this type, working on important science, engineering and business problems," he said. "You will probably never hear about those computers, because they are simply a tool; the problems that they solve and the progress they enable is the important news."

More information about Avalon is available at the following URL on the World Wide Web: http://cnls.lanl.gov/avalon

Los Alamos National Laboratory is operated by the University of California for the U.S. Department of Energy.

-30-