After the 1992 moratorium on underground testing of nuclear weapons in the US went into effect, the Department of Energy’s National Nuclear Security Administration’s (NNSA) was tasked to maintain the country’s nuclear weapon deterrent via computing simulations. As a result, Lawrence Livermore National Laboratory (LLNL) and its two sister labs at Los Alamos and Sandia became the recipients of some of the most muscular computing hardware in the world. Today these institutions are at the forefront of supercomputing expertise, both hardware and software.

Because the weapons simulation applications are always looking to achieve higher resolution, higher fidelity, and full-system modeling, there is an ongoing demand for ever-more powerful capability-class supercomputers. Today, Los Alamos houses what is ostensibly the world’s most powerful computer — Roadrunner — which clocks in at over a petaflop. In a couple of years, LLNL is slated to deploy “Sequoia,” a 20-petaflop IBM Blue Gene/Q machine, and a likely contender for the top supercomputer in 2011. Sequoia’s predecessor, “Dawn,” is a 500 teraflop Blue Gene/P machine installed earlier this year at Livermore.

But according to Mike McCoy, who heads Livermore’s Scientific Computing and Communications Department, it’s not all about these elite capability machines. He says 10 to 30 percent of the computational resources at the lab are devoted to capacity systems, that is, commodity HPC Linux clusters. The reason is simple. There is a lot of computing to be done, and time on the expensive capability systems is dear. By necessity a lot of application work has to be developed and tested on these smaller, less expensive machines as a way to contain costs.

There is also quite a bit of unclassified science work performed at the lab in the areas of climate, biology, molecular dynamics, and energy research. Some of this basic science supports the weapons programs, but the remainder is just part of the NNSA’s larger mission of furthering national security. The unclassified work also serves to nurture the lab’s scientists, and without them, there is no weapons program. In any case, the vast majority of this class of computing takes place on vanilla Linux clusters, albeit very large ones.

Today at Livermore, capacity clusters account for 404 teraflops of computing power, while the capability machines deliver 1,324 teraflops. Another 205 teraflops are available in visualization and collaboration systems. The most powerful capability system at the facility is the half-petaflop Dawn, while the largest capacity cluster is Juno, which weighs in at 167 teraflops.

Livermore has relied on a number of cluster computer vendors over the years. In 2002, the now-defunct Linux Networx installed a the MCR cluster, which delivered a 7.6 teraflops, a performance level that earned it the number three spot on the TOP500 list in June 2003. A more recent vendor is Appro, who won the Peloton contract in 2006 and then the subsequent Tri-Lab Linux Capacity Cluster (TLCC) deal, which served all three NNSA labs.

Today Lawrence Livermore appears to be grooming Dell for some major deployments. Up until last year, the only Dell machines at the lab were sitting on people’s desks. But in November 2008, the company became the cluster partner on the Hyperion project, a testbed system to be used to develop system and application software for HPC. The idea was to provide a platform for developers to build and test codes at scale before they are deployed on larger production systems. That effort has produced some early results including simulating the file system and I/O rates of the future Sequoia system using Hyperion’s InfiniBand and Ethernet SANs.

Last week, Michael Dell met with LLNL officials at Livermore to get a sense of what the NNSA is expecting from its future cluster system. The agency’s goal is to maintain at least a 1:10 performance ratio between capacity systems and capability systems. Today that means you need roughly a 100 teraflop cluster to match up with the purpose-built one-petaflop supers. With Sequoia coming online in 2011, the folks at LLNL are already thinking about clusters in the two-petaflop range. Beyond that the lab see the need for 100-teraflop commodity machines in 2018, in anticipation of capability machines hitting the exaflop mark. That means vendors need to scale today’s commodity clusters by a factor of 10 over the next 9 years.

Recently Dell installed “Coastal,” an 88.5 teraflop system that is being used by the Lawrence Livermore’s National Ignition Facility to help with fusion research. Next year, with Dell’s help, the lab will be more than doubling the performance of the 90 teraflop Hyperion system with “Sierra,” a new cluster that is spec’ed to reach 220 teraflops.

Michael Dell is hoping that’s just the beginning. From his point of view, designing systems pushing the envelope of scalability and technology dovetails nicely with the company’s other big server segments, namely web services infrastructure and cloud computing. For example, the inclusion of SSD technology to increase I/O performance in the Livermore’s Coastal cluster also turned out to be a good solution for Dell servers deployed for a Web search provider in China (presumably Baidu). He sees the demand for these super-sized machines inside and outside of HPC as two sides of the same hyperscale coin. And, he says, the technology transfer travels in both directions. “You always learn from your best customers,” says Dell.