Corps of Engineers will build its supercomputer one block at a time

Corps of Engineers will build its supercomputer one block at a time

by Bill Murray

Over the next four months, the Army Corps of Engineers' research laboratory will acquire a 512-processor supercomputer that uses fully shared memory, snapped together like Lego toys.

The $7 million Defense Department High-Performance Computing Modernization buy will include four SGI Origin 3800 supercomputers, each with 128 400-MHz processors. The computers will gradually be linked as two 256-processor systems and then a single 512-processor unit by January, via SGI's Non-Uniform Memory Architecture.

Brad Comes, director of the Major Shared Resource Center at the Army Engineer R&D Center in Vicksburg, Miss., said the first hardware will arrive at the center, one of four DOD high-performance facilities, this month.

The supercomputer will deliver at least 400 billion floating-point operations per second and has a potential peak performance of 800 GFLOPS, he said.

Clear advantages

'It will be one of the top five [SGI supercomputers] that's a single system with fully shared memory,' as distinguished from a cluster system with distributed memory, Comes said.

Creating programs for fully shared memory is easier, and it can run processes faster. Memory distributed over clusters has scalability advantages, but 'programmers have to keep track of where the data is' down to the processor level, which is difficult, he said.

Shared-memory systems until recently have been 128- to 256-processor machines, Comes said, but Army Research Lab and NASA officials are also deploying 512-processor supercomputers with shared memory.

The Vicksburg supercomputer will run the SGI Irix 6.5 operating system.

The corps will use the supercomputer to, among other things, model next-generation weapons and flight systems.

The Vicksburg site has OC-12 Synchronous Optical Network connectivity. A project partner, the University of Alaska at Fairbanks, has OC-12 or Gigabit Ethernet connectivity through the Defense Research Engineering Network, a high-speed asynchronous transfer mode network, said Frank Williams, director of the university's Arctic Region Supercomputing Center.

Fairbanks users will 'play a significant role in testing DOD applications on the large system,' Comes said. 'They will contribute to acceptance testing criteria and procedures.'

Computer Sciences Corp. is delivering, installing and testing the supercomputer.

The corps already has a 128-CPU SGI Origin 2000 supercomputer, as well as a 512-processor Cray 3TE machine and 994 processors spread over three IBM Corp. systems.

'We bought the Origin 2000 in late 1998, and it's been a good workhorse for us, but the problem sizes have grown,' Comes said. When the 512-processor Origin 3800 system is ready, the corps will be able to 'run a large problem or several smaller ones on one machine,' he said.

Stay Connected

Sign up for our newsletter.

I agree to this site's Privacy Policy.