World's largest unclassified supercomputer
The National Energy Research Scientific Computing Center has opened its newest supercomputer—a 3,328-processor IBM RS/6000 SP system—to more than 2,000 researchers at national laboratories and universities across the country
October 29—Scientists at universities and national laboratories across the country are now tapping into the power of the world's largest supercomputer dedicated to unclassified research and have reported important breakthroughs in climate research, materials science and astrophysics. The IBM SP, named "Seaborg" in honor of Berkeley Lab Nobel Laureate Glenn Seaborg, is capable of performing five trillion calculations per second (5 teraflop/s).
"Until now, this level of computing power simply has not been available to support research across a broad range of computational science," said Berkeley Lab Director Charles Shank. "As of today, however, scientists who are researching global climate change, exploring how to cut pollution from internal combustion engines, designing power sources for the future and finding new ways to treat disease have a much more powerful tool at their disposal. We fully expect this research to help shape how we live in the future."
The supercomputer is located in Berkeley Lab's new Oakland Scientific Facility in downtown Oakland. The new IBM SP boasts the computing power of more than one million desktop PCs, all able to work together to tackle some of the world's toughest scientific problems.
After thorough testing to ensure it met the rigorous demands of 24-by-7 operation, NERSC's IBM SP was opened to DOE's research community in late August. Soon afterward, scientists around the country began using its power to make important gains in studying complex problems.
"Serving up a lot of computing horsepower is only part of the computational science equation," said Horst Simon, director of the NERSC Division at Berkeley Lab. "The real measure of our success as a supercomputing center is the level of science our research community is able to achieve using our resources. We're very excited by the results already being reported and are looking forward to even greater accomplishments."
Early users of the IBM supercomputer have already reported important scientific results in astrophysics, climate research, and materials science.
The nature of the universe
A team of scientists at the University of Georgia, Berkeley Lab, and the University of Oklahoma has been using the IBM to study several astrophysics problems. Peter Hauschildt of the University of Georgia analyzed observations of very low mass stars and "brown dwarf" extrasolar planets. His advanced model grids, which would have taken several months to run on smaller computers, were completed in just four days. Not only did Hauschildt's jobs run faster, but he was able to run up to 20 models at the same time.
"The models include the absolute latest physics and detailed spherical radiation transport, which has never been done before for this type of model and for so many models at once," Hauschildt said. "The resulting model grid will be used to analyze observed spectra and to better understand the physics behind these types of stars."
Fellow researcher Eddie Baron of the University of Oklahoma ran numerous models of type IIP supernovae, which will be extremely useful for determining the extragalatic distance scale and determining the nature of the dark energy. The existence of dark energy was strongly implied when two international groups of astronomers and physicists—the Berkeley Lab-based Supernova Cosmology Project and the High-Z Supernova Search Team—discovered the accelerating expansion of the universe, using type Ia supernovae.
The unknown energy acts to overcome gravity and is thought to make up about two thirds of the density of the universe. It can be verified independently by models using type II supernovae, which Baron calculated with the new NERSC machine. Baron said the models ran up to four times faster on the new IBM than on previous supercomputers.
High resolution global climate modeling
Phil Duffy of the Climate and Carbon Cycle Modeling Group at Lawrence Livermore National Laboratory reported that his group used NERSC's IBM SP to run a global climate change simulation at the highest spatial resolution ever used for such a simulation, making the model more useful for studying regional climate change.
Global climate simulations are typically performed on a latitude-longitude grid, with grid cell sizes of about 300 kilometers. Although simulations of this type can provide useful information on continental and larger scales, they cannot provide meaningful information on regional scales, such as for the state of California, Duffy said.
"Thus, coarse-resolution global climate simulations cannot provide information on many of the most important societal impacts of climate change, such as impacts on water resource management, agriculture, human health, etc.," Duffy said.
"To do this would require simulations with much finer spatial resolution. Using NERSC's new IBM, as well as supercomputers here at Lawrence Livermore National Lab, we have been experimenting with running global climate simulations at 50 km resolution. This is finer resolution than has ever been attempted in a global climate calculation."
Compared to a typical global climate simulation, this 50-kilometer simulation has 32 times more grid cells and takes up to 200 times longer to run on a computer.
"Obviously, such a calculation could not even be attempted without access to extraordinary computational resources," Duffy said. "Our goal for the 50-kilometer global climate simulation is to evaluate how well the model simulates the present climate at this resolution. Thus far we have run about three simulated years; preliminary analysis of the results seems to indicate that the model is very robust to a large increase in spatial resolution."
Faster climate modeling
Computational scientists at the National Center for Atmospheric Research in Boulder, Colorado, are also using the NERSC IBM to run climate simulations at much higher speeds.
The goal of the team of Richard Loft, Stephen Thomas, and John Dennis is to demonstrate that a climate simulation spanning 100 years can be completed within a single day on a high-performance computer such as the IBM SP. This represents a 10-fold increase in speed over the current rate "and would represent a major advance in geophysical fluid dynamics," according to the group.
Using 2,048 processors on the NERSC IBM supercomputer, the team demonstrated that the dynamical core of an atmospheric general circulation model (GCM) can be integrated at a rate of 130 years per day. Future work will include parameterizations of physical processes such as precipitation and solar radiation.
Running such models at high speeds is critical in determining the sensitivity of a climate model to the variation of atmospheric constituents in a comprehensive simulation of the Earth's climate. A complete model of the Earth's climate system would also couple the atmospheric component to ocean, ice and land surface models. Such coupled models also require massive computing power such as that provided by NERSC's IBM system.
Better understanding of magnetic forces
As part of the extensive testing of the IBM SP system, a DOE research team of scientists from Oak Ridge National Laboratory, Lawrence Berkeley National Laboratory, and the Pittsburgh Supercomputing Center used the supercomputer to perform first-principles spin dynamics simulations of the magnetic structure of iron-manganese/cobalt interfaces.
These large-scale quantum mechanical simulations, involving 2016-atom super-cell models, reveal details of the orientational configuration of the magnetic moments at the interface that are unobtainable by any other means. This work is of fundamental importance in improving magnetic multi-layer computer storage and read head devices. Using 2,176 processors on the IBM SP, the team achieved a maximum execution rate of 2.46 teraflop/s, one of the highest levels ever for a code producing significant scientific results.—by Jon Bashor