Lab Avoids Storage OverloadLab Avoids Storage Overload

Any business grappling with scalability issues should visit a federal laboratory sometime soon. Engineers at Los Alamos National Laboratory process and store unprecedented amounts of information, and they can't afford to slow down. In two years, their requirements likely will increase more than tenfold.

For Los Alamos National Lab, dealing in teraflops--a trillion floating point operations per second--is common. The lab supports 20 teraflops of military weapons-testing simulations that process information at speeds of 20 Gbytes per second.

"A single calculation might run for a year, and no single machine could stay up that long," says Gary Grider, group leader for high-performance systems integration at Los Alamos. That's why the lab uses a Linux cluster with 12,000 processors to process many terabytes of data in very short times. "Two years from now, tests will be 300 teraflops, and we'll need 100,000 storage disks for performance. We continue to scale bigger and bigger," Grider says.

To help achieve this scalability, Los Alamos uses object-based storage from Panasas Inc. The Panasas Active File System 2.0, in support of the Linux cluster, lets the lab reach the performance levels it needs at a current rate of about a gigabyte per second per teraflop. Parallel data access provides breakthrough performance, file locking lets multiple applications function well next to each other, and load balancing helps with business continuity.

"The object model lends itself to the parallel processing we need," Grider says. Panasas was the only vendor to respond to the laboratory's request for proposals, which included support for 10 teraflops and the gigabyte per second speed, he adds.

The vendor and the laboratory are a good match, one industry analyst says. "Other vendors are starting on Linux clusters, but Panasas is driving innovation in developing scalable data solutions," says Addison Snell, an analyst at IT market-research firm IDC. "It makes sense for Los Alamos to take this approach with its scalability needs."