Sandia Blasts Off Blade Cluster: Page 2 of 3

The initial cluster setup offered a peak performance of just over half a Teraflop (trillions of calculations per second), Oefelein says. It wasn't enough. The Livermore, Calif.-based CRF is now doubling that capacity with the addition of 72 more blades. With a new peak performance of 1 Teraflop, the CRF is planning to do some serious number crunching.

Prior to deploying the cluster, the facility relied on an eight-year-old Origin server from Silicon Graphics Inc. (SGI) (NYSE: SGI). Although this fit the bill for awhile, the system had reached “the end of its life,” according to Oefelein, and, like a growing number of other high-performance computing sites, the CRF opted for a cluster (see Buffalo Cluster's a Grid Cornerstone and Panasas Powers Stanford Research).

The new setup is giving CRF more bang for its limited buck -- about $800,000, according to Oefelein, who also cites scaleability and ease of use as major benefits of the cluster. Using a modified version of Linux from Penguin’s subsidiary Scyld Software to actually control the blades helps, too. “Linux is the operating system that we all know,” he says. “Some of the guys here can rebuild the kernels while they are eating their lunch.”

The CRF cluster uses InfiniBand as its interconnect fabric, which speeds up data transfer by “orders of magnitude,” according to Oefelein, although he is unable to put an exact figure on this. Storage comes from nStor Technologies Inc.'s (Amex: NSO) RAID arrays, which provide around 13 Terabytes of disk space.

The cluster would have made it onto the June 2004 list of the Top500 Supercomputing Sites. But things have moved on significantly in the last year, and the lowest-placed site on the most recent list performed at 2.3 Teraflops(see Top Supercomputers Revealed).