Exploring the most advanced computing systems available, Ames Lab researchers are working to make today's complex scientific questions more approachable.
"Clarifying uncertainties such as global climate change requires a radical departure from traditional computing," says John Gustafson, Ames Lab computational scientist at the Scalable Computing Laboratory (SCL). A cooperative effort of Ames Lab and Iowa State University (ISU), the SCL is one of the most powerful and innovative supercomputer facilities in the country.
The focus of the SCL is parallel processing rather than the traditional serial processing of most computers. Parallel computers have multiple processors simultaneously working on different parts of one problem while serial computers tackle a problem one step at a time, generally with one processor. "The difference can be compared to the difference between managing one worker or managing a group of people working together on a project," explains Gustafson. The group of people is able to tackle a more complex project, although managing those people may take more planning and organization.
While parallel computers provide the fastest calculations available and lower costs substantially for those calculations, the complex architecture of parallel processing and an inherent resistance to change have limited commercial development. "With trillions of dollars invested in serial hardware, software and mind-set, it will take the commitment of progressive non-profit institutions such as Ames Lab to tap the potential of this new technology for scientific investigation," says Gustafson. He explains that Ames Lab and ISU administrators and researchers have an innovative attitude toward computing and a willingness to overcome the hurdles which come with breaking new ground. The Lab's progressive enthusiasm drew Gustafson to Ames two years ago from Sandia National Laboratory in Albuquerque. At Sandia, Gustafson and his colleagues won an R&D 100 award and the Gordon Bell Award for the first clear demonstration that parallel processors can work efficiently on a single problem and surpass the speed of conventional supercomputers.
"The people, priorities and timing at Ames Lab are right for it to play a vital role as the nation focuses on high performance computing as a key technology for enhancing U.S. competitiveness," says Gary Johnson of the Department of Energy's (DOE) scientific computing office in Washington, D.C. He says evidence is growing that parallel computing will be the way of the future. Until the breakthrough of Gustafson and his Sandia colleagues, prevailing theory held that parallel processing would offer only limited performance gains. Now, dozens of computer corporations, including serial computing leaders like Cray Research, Inc., are committing themselves to the parallel computing market.
The new SCL is evidence of the priority Ames Lab and ISU are giving this technology. An outstanding example of true collaboration, the facility is jointly funded by DOE and ISU's Institute for Physical Research and Technology.
The facility includes a MasPar unit with 8,192 processors and an nCUBE model with 64 processors. The Ames Lab MasPar, one of the fastest computers in the world, can complete 30 billion operations a second. The SCL was one of the first MasPar installations in the world, acquiring the fourth machine available. About 20 of these massively parallel machines are in use now.
"We are committed to buying and keeping up-to-date with the most aggressive, commercially available designs and breaking free of traditional programming methods to achieve the highest performance and improved cost-effectiveness," says Gustafson. The concern in choosing the hardware is to provide a unique collection of machines offering a variety of approaches to parallel computing but always maintaining scalability or the capacity to increase the processing power for future needs. Scalability has been lacking in most serial processing systems, quickly leaving them obsolete.
A top priority is figuring out new ways to use the intricate machines more intelligently and efficiently, thus making them more useful to scientists of diverse disciplines. The SCL encourages interdisciplinary research on the systems and the computational scientists work with individual researchers in using these tools to ask increasingly complex questions. Scientists at Ames Lab, ISU and other government laboratories are already taking advantage of the SCL's unique computing capabilities. Current University and Lab projects include atom scattering simulation, digital image processing and global climate modeling.
"Ames Lab is making a significant effort toward more intelligent use of these machines and it provides a unique opportunity for educating the computational scientists of the future," says Johnson. A priority for both the Lab and ISU, students from the high school through graduate level are learning the intricacies of this new technology. A group of undergraduate and high school students with ISU's Women in Science and Engineering Summer Internship Program provides an interesting example. The students learned to write programs for the machines and have just completed manuals to help new users learn the basics.
A good illustration of the benefits possible from Ames Lab's unique relationship with ISU, the new Scalable Computing Laboratory will help expand the understanding of this vital new technology.
Comparing various computers is a difficult challenge, but Ames Lab researchers have found that a non-traditional approach provides improved assessment. While previously developed benchmarks fix the size of the problem to be run during the evaluation, SLALOM fixes the time of the comparison test.
"Consider the comparison of a jet and a tricycle," says John Gustafson, one of SLALOM's developers. In a 100-mile race, the tricycle might not finish, but in a 100-foot race, the tricycle would win." Clearly, the better question for a fair evaluation is how far can each go in a set amount of time," he says to explain the theory of the fixed-time benchmark.
SLALOM, which stands for Scalable, Language-independent, Ames Laboratory, One-minute Measurement, is the first Scalable supercomputer benchmark. Adaptable to future computer technology and unprejudiced toward any machine or programming environment, SLALOM is useful on scalar, vector and parallel machines of all kinds. It can compare an Apple Macintosh personal computer, a CRAY supercomputer and a MasPar parallel processor, doing justice to all three.
"This sort of comparison is particularly important if computer buyers are to explore new approaches in computing and take advantage of the latest technological advances," says Jeff Kalb, president of MasPar, a parallel computer manufacturing company.
While vendors and other outside organizations are welcome to run SLALOM on their machines and many already have, the SLALOM report always identifies who ran the benchmark. Upholding high standards of credibility and serving as an independent, unbiased clearinghouse for these computer evaluations, Ames Lab maintains all SLALOM performance data.
Published by Inquiry, winter 1991
Contact:
John Gustafson
+
Gus@SCL.AmesLab.gov
Pages prepared by Maria E. Blanco.