Los Alamos has fastest supercomputer

The Laboratory on Nov. 10 claimed bragging rights for the world's fastest supercomputer.

The computer, known as Blue Mountain and built by Silicon Graphics Inc., ran Linpack, the industry's standard test, at 1.608 trillion calculations per second, or teraOps.

The Lab and SGI also announced that the Laboratory has installed the most powerful visualization system to enable scientists to make sense of the unprecedented amounts of data from Blue Mountain.

"With Blue Mountain and SGI's visualization engines, we are beginning a new era in computing and in science, moving the stewardship of our nation's nuclear weapons from its 50-year foundation in nuclear testing to one based in science and simulation," said Steve Younger, associate Laboratory director for nuclear weapons.

"These computing tools are the first step toward what we need to maintain confidence in an aging stockpile and our commitment to the Comprehensive Test Ban Treaty," Younger added.

The Laboratory bought Blue Mountain from SGI through the Accelerated Strategic Computing Initiative, or ASCI, a program funded through the Department of Energy's office of Defense Programs. One of ASCI's goals is to build, over the next five years, an even more powerful supercomputer that can execute 100 trillion calculations in a second.

The Laboratory last week ran other tests on Blue Mountain, which was completely installed and running at the Lab on Oct. 30. Those tests include a benchmark code called simplified Piecewise Parabolic Method, or sPPM, which simulates shocks and turbulence in fluids and is more similar to real weapons codes than the Linpack test.

At the heart of Blue Mountain are 48 Silicon Graphics/Cray Origin2000 servers containing a total of 6,144 processors, with projected peak performance of 3.072 teraOps.

More importantly, computing of nuclear weapons simulations done on parts of Blue Mountain over the past two years indicates that the machine will obtain greater sustained performance compared to previous computers. High sustained performance -- not theoretical peak performance -- is crucial because large-scale weapon simulations typically run across the entire supercomputer, stressing all system components.

"The truly awesome speed and power of Blue Mountain in these tests is a tribute to months of hard work by scientists, engineers and technicians from Los Alamos and SGI," Younger said. "But the real benefit to the nation comes now, when we can begin harnessing the computer's power to make accurate predictions about the future of the nuclear stockpile.

"In solving these urgent national security problems, we also are creating a scientific revolution in how human beings use machines to help them think," Younger said. "Blue Mountain can do more calculations faster than other computers. But once you have trillions of bits of information, you need the most powerful visualization engines to extract knowledge from that data and see it in three dimensions. And finally, you need new, powerful computer codes and operating environments to give scientists a more accurate representation of reality."

Blue Mountain is organized into 48, 128-processor shared memory multi-processors, or SMPs. The system is designed so the cluster of 48 SMPs -- all commercially available servers -- behave like a single computer.

These 48 SMPs can communicate with each other at world-record sustained speeds in excess of 650 gigabits a second. This high bandwidth is because Blue Mountain's SMPs are interconnected by HIPPI-800 switches. HIPPI, or high-performance parallel interface, is a supercomputer communications link that Los Alamos helped make a global standard. Early next year, Blue Mountain will become the first computer to run the new HIPPI standard, called HIPPI-6400, which is eight times faster than standard interconnects.

Storage systems also will break records. Blue Mountain will have more than 76 trillion bytes of state-of-the-art fiber channel disk, making it one of the largest "disk farms" ever assembled for a single computer.

The Lab and SGI also have completed installation of the largest system in the world for visualization of computer output, comprised of 16 of SGI's Infinite Reality graphics engines. This dedicated visualization system permits scientists to view and refine information sets as large as one billion cells in a single interactive picture. The visualization system is especially valuable because it is an intrinsic part of Blue Mountain, not a separate unit.

"Blue Mountain represents a new era in supercomputing," Younger said. "Instead of ordering one-of-a-kind machines from a highly specialized market, we're working with SGI to create the most powerful computer on the planet based on off-the-shelf technologies that are of direct value to the half-trillion-dollar-a-year commercial computing market."

In September, using just one-sixth of Blue Mountain, the Laboratory's weapons scientists were able to run a weapon simulation that may help solve a decades-old mystery raised by a past underground nuclear test. The simulation ran across 30 million zones, the largest weapons simulation of its class.

Through ASCI, the DOE initially is providing 3-teraOps-scale supercomputers at two of its nuclear weapons laboratories: Los Alamos and Lawrence Livermore, and a 1.8 teraOps system at Sandia. Program plans include a 10-teraOps supercomputer by 2000, a 30-teraOps machine by 2001 and a 100-teraOps supercomputer by 2003-4. Next year, DOE plans to award a contract for the 30-TeraOps ASCI supercomputer at Los Alamos following competitive bidding.

In addition to its extraordinary speed, Blue Mountain is giving weapons scientists a much better computer picture of the weapons in their care. During 1999, Blue Mountain should execute 80 million trillion operations over the course of thousands of ASCI simulations focused on key weapons issues. This is roughly 10 times more computing in one year than all the calculations executed in support of the U.S. stockpile from the Manhattan Project through 1992, the last year of underground testing.

The first supercomputer at the Laboratory, "MANIAC," was built in the early 1950s based on the ideas of computer genius John von Neumann. The nuclear weapons program at the Lab was the birthplace of high-performance computing; the first model of nearly every supercomputer was sold to the Laboratory, including the first Cray supercomputer ever manufactured, which was installed at the Laboratory in 1976.

Advances in supercomputing at Lab and its sister laboratories have driven computing to the modern era in which personal computers have more power than the machines used in the 1950s to design the first thermonuclear weapons.

Through the DOE's Office of Energy Research, the Laboratory and SGI are building a second supercomputer, Nirvana Blue, which will be capable of one teraOps of nonweapons computing. Academic and industrial supercomputer users are working with the Lab to develop and improve predictive simulations on Nirvana Blue. Because Nirvana Blue and Mountain Blue share an identical architecture, these unclassified advances will be invaluable in helping to improve the state of the art in weapons simulations.

Nirvana Blue, which was funded by DOE's Office of Energy Research, currently has 768 of its 2,048 Origin 2000 processors installed. Installation should be complete by March 1999.

The Laboratory is coupling its advances in nuclear weapons simulations with new predictive capabilities in such scientific areas as the modeling of global climate, wildfire, disease epidemics and other crises, and transportation, electric power and other infrastructure.

Laboratory computer scientists are working with the oil industry to analyze seismic data to locate new wells and increase the amount of oil recovered from existing wells. The automobile industry uses computer codes invented at the Lab to model and improve the efficiency of internal combustion engines. And Laboratory researchers, with partners in academia are in the forefront of studies of the world's climate and how it may be affected by human activity.

--Jim Danneskiold