Air Force IT pumps up

To help meet its seemingly insatiable demand for supercomputing power, the Defense Department will take delivery of a new supercomputer, to be installed next month at the Air Force's Aeronautical Systems Center's Major Shared Resource Center, located at Wright-Patterson Air Force Base, Ohio.

The acquistion is the second this year for the center.

'There is much more demand in the DOD environment than all the resources that the centers in high performance computing program have put together,' said Steve Wourms, director of the center. 'On top of that, researchers are leveraging these much-improved machines to do much higher resolutions or much finer grids.'


For this contract, Hewlett-Packard Co. will install a 2,048-processor HP Cluster Platform 4000 system, which is expected to be able to execute 10 trillion floating-point operations per second (10 TFLOPS). The contract, which includes a number of smaller systems as well, is for $8.6 million.

The Defense Department's High Performance Computing Modernization Program Office oversees this effort to supply supercomputing power, as a shared resource, to other Defense research, development, test and evaluation communities. The modernization program spends between $35 million and $60 million a year on new systems, according to Cray Henry, a manager of the DOD program who spoke at a meeting of the Beowulf Users Group in Washington. About 4,500 military personnel use the systems for approximately 560 projects.

The program has designated four Defense organizations, called Major Shared Resource Centers, to run the supercomputers. In addition to Wright-Patterson, the Naval Oceanographic Office at the Stennis Space Center, Miss., the Army Engineer Research and Development Center in Vicksburg, Miss., and the Army Research Laboratory at the Aberdeen Proving Ground, Md., all run MSRCs.

The program's technology refresh program usually acquires new systems for two of the four MSRCs a year. This year, Wright-Patterson and the Army Engineer Research and Development Center got the upgrades. In March, the Vicksburg facility received an 18-TFLOP machine from Cray Inc. of Seattle. Next year, the two other facilities will get upgrades.

Wright-Patterson's computer, like others in this program, will run those jobs that would overwhelm departmental computing resources, said Jeff Graham, deputy director of the ASC MSRC. The jobs range from designing weapons and undertaking large-scale threat re- duction studies to providing high-quality simulations of fluids and structures. The ASC MSRC provides not only the raw compute cycles, but makes a number of Ph.D.'s available to help with problems such as code parallelization and visualization services.

The HP system will be installed in September, making it the second new system for the center this year, according to Graham. In March, the center took delivery, under another bundled contract worth about $13 million, of a 2,048-processor SGI Altix System. Nicknamed Eagle, that machine has shown a peak performance of over 11.636 TFLOPS.

Selection Process

The ASC MSRC computers were procured through an exhaustive open competitive process, one involving a wide variety of tests to determine the best configuration, Graham said. 'We need a high-performance computer to analyze the high-performance computer we're considering buying,' Graham said.

The program office evaluates how well a potential system performs by having the vendor run a sample of Defense applications on a prototype of the machine, Henry said. The resulting performance metrics are weighed against more standard benchmarks, I/O speeds, operating system performance, memory, and network and CPU speeds of individual systems.

All these performance metrics are then weighed against the price of each system. The center also looks at qualitative factors, such how comfortable the administrators are working on a particular system. Usually, each round of procurement takes about nine months to complete, Henry said.

Strength in Diversity

Each system purchased has slightly different performance characteristics, Graham said. The variation will ensure that the program has the widest variety of options for providing services.

'We'll buy a portfolio of systems that best meet all of our benchmark codes,' Graham said. 'They are configured slightly differently. Some benchmark codes run better on certain systems.

Using a shared-memory design called NUMAflex, the SGI machine would be particularly useful for those jobs that take up an immense amount of working memory. Each compute node incorporates 500 processors, said Ralph McEldowney, chief of the ASC MSRC advanced technologies branch.

In contrast, the HP system is a traditional cluster, with each node using only two processors. This cluster would work more readily with programs that can be split up and distributed across the memory banks of multiple machines.

'The two systems are architecturally very different. The systems meet different requirements because they run different jobs,' McEldowney said.

With these purchases, Wright-Patterson now has four major supercomputing systems, capable of pumping out a maximum of 27 TFLOPs, more than five times the 5 TFLOPs available last year. The boost is sorely needed, Graham said, as large jobs can now take up 90 percent of a system.

'It is a big leap forward for us to give our users much more capability in terms of raw power, memory and disk space,' Graham said.

About the Author

Joab Jackson is the senior technology editor for Government Computer News.

Stay Connected

Sign up for our newsletter.

I agree to this site's Privacy Policy.