Portland, Ore. Oak Ridge National Laboratory's Center for Computational Science plans to use a $25 million cash infusion from the U.S. Department of Energy to help regain bragging rights to the global supercomputer speed record for the United States. The award, announced last week by U.S. Secretary of Energy Spencer Abraham, is part of a $300 million modernization program at ORNL. The goal of the five-year program is a supercomputing capability of a sustained 50 teraflops and a peak capacity exceeding 250 teraflops.
Abraham said the National Leadership Computing Facility (NLCF; Oak Ridge, Tenn.) will pool the computational resources of two national labs ORNL and Argonne to create the world's fastest supercomputer. He predicted faster machines will yield breakthroughs in such areas as weather prediction, biological modeling, battlefield visualization and nanoscale materials discovery.
"There are two major reasons why our efforts are important to the nation," said Thomas Zacharia, ORNL associate laboratory director for Computing and Computational Sciences. "First, as we push the envelope building these very, very high-speed machines, we will develop technologies that then percolate down to enterprise servers and eventually even personal computers.
"Second, the engineering and scientific breakthroughs that result will affect us in our daily lives from climate science and energy efficiency to advances in nano technology, the discovery of new materials, the life sciences, biotechnology and new drug discovery. These all impact our humanity as a nation. We need to have the technology to enable these breakthroughs, so that our country can maintain its leadership role and economic vitality."
The new computing capacity will be rolled out over the next five years. A supercomputer installed at the Earth Simulator Center in Japan currently holds the global speed record, at nearly 40 teraflops. Even with an upgrade by year's end, the NLCF will offer, by contrast, only a total of 20 teraflops from its single Cray X1 computer at ORNL. But by next year NLCF should reclaim the crown for the States by partnering with Argonne National Laboratory to offer more than 45 teraflops, albeit via three machines: the 20-Tflops ORNL Cray X1 computer; another 20-Tflops Cray, this one a Red Storm-based processor; and an IBM Corp. Blue Gene from Argonne that will provide the 5 remaining Tflops.
The NLCF's use of Cray and IBM supercomputers stands in contrast to the clusters of microprocessors used in Japan's machine. The two Crays in the NLCF program, one scalar and one vector, will be tightly coupled to run a single application at 40 teraflops. In contrast, the Earth Simulator runs a dedicated application over 640 NEC Corp. microprocessors connected through a single-stage crossbar switch with memory-mapped access to eight vector-based microprocessors.
In 2006, the NLCF will add a 100-Tflops Cray that will break the single-machine speed record and combine with the three other supercomputers to provide 145 Tflops. By the end of the five-year program, the 100-teraflops Cray is expected to be upgraded to 250 teraflops, bringing the total computing capacity to 295 teraflops.
"While it is in some aspects about the computers themselves, the biggest opportunity will be the new discoveries that can be made," said Zacharia.
Argonne and ORNL also plan to house a next-generation suite of modeling software at the NLCF to facilitate scientific-computer users' access to the engineering and scientific breakthroughs that the upgraded supercomputing capability is expected to enable.
The NLCF occupies a new, 170,000-square-foot facility, with 40,000 square feet dedicated to a vast computer room that the Tennessee Valley Authority has wired with a 12-megawatt power supply. The remaining square footage houses a staff of 400, including "mathematicians, computer scientists, computational scientists, networking experts, visualization experts, user-support personnel, system administrators and of course a few managers," said Zacharia.
Separately, IBM announced last week that it has opened a second cluster-supercomputing center that takes advantage of hundreds of microprocessors but interconnects them using conventional networking approaches that are customized for specific applications. The Deep Computing Capacity On Demand Center in Montpellier, France, houses 896 processors that can be clustered in a variety of user-specified ways to solve specific application problems.
IBM's first supercomputer center, in Poughkeepsie N.Y., lets users network more than 2,300 processors.
"We are starting off with a dozen eServer pSeries 690s, each with 32 CPUs, as well as xSeries servers with a total of 512 Intel Xeon processors," said Mark Solomon, director of Deep Computing Capacity On Demand. "Our users can configure them to be completely separate so their data is always secure."
The IBM supercomputer centers offer users the ability to "pay as you go" for supercomputing capacity, enabling what-if problems to add computing capacity as a design variable, according to IBM.