DOD upgrades super tech

The Army Corps of Engineers Waterways Experiment Station (CEWES) late last month announced it had awarded $25 million in contracts to upgrade its supercomputers and to take advantage of some of the newest and most powerful technology available.

The upgrades, which also are occurring at other Defense Department laboratories participating in the DOD High-Performance Computing Modernization Program, will provide Vicksburg, Miss.-based CEWES with a total peak processing power of 1.4 trillion floating-point operations per second (teraflops). The lab will use the new systems for some of its largest computing problems, including research on lasers for DOD's ballistic missile defense program and research on making tanks more efficient at clearing land mines.

The systems that CEWES purchased include a next-generation RS/6000 SP supercomputer from IBM. The supercomputer is designed to let researchers employ "mixed-mode" programming, combining software written according to different programming models in the same application. "Now they have to make a choice [and] do one or another," said Bradley Comes, director of the CEWES Major Shared Resource Center. "There are performance trade-offs."

CEWES will be among the first customers to install an SP built with IBM's new Power 3 microprocessors, which the company announced last month. Bill Kneisly, IBM client manager for CEWES, said the Power 3 architecture makes mixed-mode programming possible by enabling different functions to be dedicated to specific clusters of processors.

In addition, CEWES upgraded its Silicon Graphics Inc. (SGI)/Cray Research T3E computer from a 336-processor, 450 MHz system to one with 544 processors running at 600 MHz.

Scalable vs. Vector

Each of these systems, plus several others already installed at CEWES, is based on scalable parallel technology, in which computers are programmed to distribute pieces of an application among many processors. With the upgrades, CEWES is ready to retire its last supercomputer based on vector technology, a Cray C90.

Use of vector technology, in which the computer processes one part of an application at a time, has declined steadily in the past few years, largely because scalable parallel machines built from commodity parts are less expensive. U.S. supercomputer vendors are abandoning the architecture, although some Japanese companies still use it.

"It is no longer really affordable to have performance that depends upon the ultimate screaming speed of the high-performance vector machines," said Richard Partridge, program director for parallel open-systems hardware at D.H. Brown & Associates.

DOD has been following this trend. For example, the Aeronautical Systems Center at Wright-Patterson Air Force Base, Ohio, is decommissioning its Cray C90 this month. Nevertheless, some labs still require vector-based systems.

In January, the Naval Oceanographic Office (Navo), Stennis Space Center, Miss., bought a Cray Scalable Vector (SV) 1 system from SGI so that the office could continue to use its vector-based software. SGI introduced the SV1 last year to satisfy demands from users who still want vector machines at a lower cost, but the system is intended as an interim step as customers migrate to newer designs.

Navo also upgraded scalable parallel systems, adding processors to its Cray T3E system for a total of 816 and buying a 128-processor SGI Origin 2000 system that it plans to integrate with another already on site. Nevertheless, a Navo spokesman said, "vector is required."

"For those applications that had been tuned and optimized to be screamingly fast on vector machines, vector machines are the product of choice," Partridge said. Very complex applications, such as ones that support unique Defense research, may be hard to migrate.