US Army Purchases 5 Linux Supercomputers for TI-06 Upgrades
The ability to use open-source operating systems like Linux with “clusters” of computing hardware that include many commodity components has really changed the supercomputing landscape. Complex physics-based problems that are three dimensional and time-dependent are especially difficult, and benefit greatly from the additional computing power per dollar that’s rapidly becoming available. This has had a significant impact in sectors like movies and TV, energy, and pharmaceuticals – not to mention the defense industry.
The DoD’s High Performance Computing Modernization Program (HPCMP) recently placed an order with Linux Networx for 5 supercomputers as part of the Technology Insertion 2006 (TI-06) initiative. They are part of a broader effort to modernize the US DoD’s computing capabilities, by providing the supercomputer services, high-speed network communications, and expertise for U.S. Defense laboratories. The recipients, and the type and capabilities of the systems they received, are detailed below:
The US Army Research Laboratory’s Major Shared Resource Center (ARL-MSRC) hosted the first large computer (ENIAC) in 1946, and sits at Aberdeen Proving Ground in Hartford County, MD. They received 4 of the supercomputers: 3 of Linux Networx high-end Advanced Technology Clusters (ATC) and a mid-tier LS-1. The most powerful ATC is expected to be ranked in the top 20 of the world’s most powerful computer systems, and the procurement increases the Major Shared Resource Center’s computing capability from 36 trillion floating-point operations (teraflops/ TFLOPS) to over 80 TFLOPS, making it one of the largest computing centers in the US Department of Defense. See corporate release.
Dugway Proving Ground in Utah received the remaining system, also a mid-tier LS-1 system. The Department of Defense has designated Dugway as the primary chemical and biological defense testing center under the Reliance Program. Testers here determine the reliability and survivability of all types of military equipment in a chemical or biological environment.
While the ARL-MSRC isn’t disclosing the purchase price, Computerworld reports that IDC analyst Earl Joseph places the average cost for a cluster at about $2,000 per processor compared with $12,000 per processor for a traditional RISC-based, symmetric multiprocessing shared-memory systems. Of the total $9.1 billion high-performance computing market last year, IDC believes that clusters accounted for about half of the sales.
Unsurprisingly, The Linux Supercomputing Company recently announced record orders and growth in 2005. The company achieved three consecutive quarters of record growth in orders in Q2, Q3 and Q4, and finished the year with a bookings backlog that increased 300% over 2004. During 2005, one of every three Linux Networx supercomputing systems was sold to a new customer, expanding the company’s customer base to over 165 leading supercomputing customers. New customers during 2005 included Audi, BMW, Caterpillar, DaimlerChrysler, EADS Eurocopter, Glaxo SmithKline, Harley-Davidson, Magna Intier and Magna Seating, Motorola, and Thomson Industries.
Geek Sidebar: System Details
The largest ATC system in this order is expected to be ranked in the top 20 of the world’s most powerful computer systems, increasing the Army Research Laboratory’s computational capability by more than 28.7 TFLOPs. It is a 1,122-node supercomputer with 4,488 mid-voltage 3.2 GHz Intel Dempsey cores to do the computing. The system will also have 112 more 3.46 GHz cores (28 nodes) for log-in, storage and administration, with 9.4 TB (9,400 GB) of memory and 260 TB of raw disk space. All nodes will communicate via a 4X DDR (20 Gbps) Infiniband network with 10 Gigabit Ethernet uplink capability. The 3.2 GHz mid voltage cores were selected for the compute nodes due to their thermal efficiency – i.e. their tendency not to overheat when you pack 4,600 of them into a small area.
Note that $4,600 x IDC’s $2,000/core estimate = $9.2 million.
A second ATC system may also crack the top 20, at least for a little while. This 842-node ATC will use 3,368 mid-voltage 3.2 GHz Intel Dempsey cores for computation, increasing ARL-MSRC’s computational capability by more than 21.5 TFLOPS.
Compare these two systems with DID’s March 2005 coverage of the USAF’s “Eagle” supercomputer at Wright-Patterson AFB, OH – a 2,048 processor SGI Altix system that offers 11.636 TFLOPS, purchased as part of the TI-05 initiative.
The third ATC system is merely a 68-core test and development system with 7 TB of disks. It will be delivered in advance of the other two systems, and probably still packs more GB in its RAM than you and two buddies have available on all of your hard drives, combined.
The mid-tier LS-1 systems at ARL-MSRC and Dugway use AMD Opteron 200 family cores; and can be fitted for up to 128 nodes/ 512 cores with up to 19.2 GFLOPS theoretical peak performance per node, if using the AMD Opteron 280 Dual Core. While this gives current LS-1 systems a theoretical maximum of 2.457 TFLOPS, in real-world operation the performance limit would be considerably lower.