ASTRONOMY

Covering the fiscal years of 2002 through 2006, developing the plan was a "major effort," when DOE accepted it, the department mentioned NERSC's "excellent staff" and said "they look forward for the next five years to support NERSC and Berkeley Lab to continue providing superior service," NERSC Director Horst Simon said in a recent exclusive interview. "Given that there are uncertainties all over the world, this is as sure as a guarantee that you can get," he said. "From that point of view, I can just sit here and I can happily smile. None of the vendors here has that much certainty in their future." The five-year plan outlines the things that NERSC plans to do to continue to provide the DOE's Office of Science with computational and storage resources, Simon explained. "In the past, NERSC always said--and this is a model that all the supercomputer centers follow-- that we have big iron hardware and the people to help users," he said. "We extended that to include a strategy which would help large collaborations in computational science--science challenge teams--to make more efficient use of high-end machines. "That's important because I think there's a big trend toward large computational science teams-- away from the single, principal-investigator, model," he added. "We want to do this but we want to continue, of course, to support the traditional user base as well." As part of the plan, NERSC put forward the idea to place the center as a computational and storage resource on the DOE Science Grid. "I think the big differentiator is that we actually perceive this [grid technology] as something that will ultimately be part of our operational environment," Simon said. "If you talk about grids today, they are in many ways, prototypes or pilot applications." For grids to become production systems, however, Simon said that the technology must be made more mature, implemented to be more stable, and then an operational model developed which provides for 24/7 support. "We put this in the five-year plan. It is something which probably takes five years to get [to]," he said. The idea of computing and data grids seems "far out," but Simon compared their emergence to the events a decade ago at the Supercomputing 91 conference, where attendees saw the first major wave of highly parallel machines, like those from Intel and Thinking Machines, make a big splash. "The age of massively parallel computing really arrived at that conference," Simon said. "But I think it took from ‘91 to about ‘96 or ‘97 for some machines like the [Cray] T3E to arrive, which really had all the production tools [and] were robust and reliable for full-production environments. So five years in terms of maturation time is not that long." As grids become pervasive, they will open NERSC to "a number of more applications which in the past have not been served very well by end high-end supercomputing centers," he said. For example, Simon noted that in the area of high-energy physics, "the fact that you can connect this cluster computing and storage resources directly to the experiment [with grids] will probably lead to a new quality of scientific discovery." Simon is also very pleased that NERSC's new IBM SP system, called Seaborg, has moved to production and has been upgraded to a full 5-teraflops peak performance level. In its early usage phase, the system achieved "some very good scientific results," he said. "More results are coming in," he added. "It just turns out to be a real great tool for computational science." Asked about the levels of sustained performance the new computer is producing, Simon called it "tricky question." On applications that are communication-intensive and require large amounts of bandwidth, "we get on the order of 7 percent of peak." That's on the low end--on some other calculations, the NERSC IBM computer achieved more than 60 percent of peak. "As usual, your actual mileage will vary," Simon quipped. "The result is somewhere between these two extremes." Simon said the "ultimate measure" of success in high-performance computing is not in the numerical, quantitative measures, but if the machine is producing "a quality [scientific] product." "In hindsight, we can look at some of the accomplishments of the last two or three years and say, ‘These were real scientific breakthroughs which were enabled by computational technology,'" he said. Simon cited a couple of examples, including a major breakthrough in astrophysics--the discovery of most distant supernova. That advance, he said, was based on data analysis at NERSC. "I know that we're very much under scrutiny to deliver quantitative measurements of results," he said. "But I think that what we really have to do is communicate the science story so that the public can understand the good and positive things that come out of computational science." Whether an application uses 7 or 70 percent of a machine, "if you get the result out on time and continue insight to scientific discovery, then it doesn't matter; the money is well-spent," Simon said. NERSC officials, Simon said, have "seen really a large number of users being very productive. "You want to increase the efficiency and use of the machine if you can, but we know that there are just some applications which don't lend themselves to higher efficiency," he said. ---------- Scott Nance is the editor of New Technology Week and is based in Washington, DC. ---------- Supercomputing Online thanks Horst Simon for his time and insights. ----------