Linux On Steroids

You can salivate over the fastest computers in the world by browsing http://www.top500.org, or you can build your own. Sandia National Laboratories, which conducts research under the auspices of the Department of Energy, has released Cplant open-source software for building Linux supercomputing clusters. The 50 Mbytes of source code is available free to those who want to build a supercomputer using off-the-shelf hardware. Cplant is designed to link PCs running Linux to create highly scalable, parallel computing clusters typically used for intensive number-crunching required by activities such as seismic research and economic forecasting.

Building your own Linux cluster capable of supercomputer-class computing still will cost thousands, but by using conventional PCs running Linux, a Cplant cluster will be several orders of magnitude cheaper than a Sun E10000 Starfire, a Fujitsu VPP5000, or any other name-brand supercomputer or proprietary clustering system.

"Our systems are roughly four times cheaper for the same Linpack number," says Neil Pundit, department manager of scalable computer systems at Sandia, in Albuquerque, N.M. Linpack is a benchmark program widely used to rate the basic computational performance of supercomputers. A 1,000-node Cplant cluster ranked 31st last year on the http://www.top500.org site with a Linpack number of 512 gigaflops. The Cplant software can be downloaded at http://www.sandia.gov.

Sandia's Linux cluster, Antarctica, is one of the largest and most powerful in the world, with 1,500 rack-mounted Compaq 1U servers, each with a 466-MHz Alpha CPU and 256 Mbytes of RAM.

The machines are interconnected using Myricom Inc.'s Myrinet gigabit clustering hardware, which is widely used for clustering and is sold by 18 of the 20 largest computer companies. The Cplant software requires you to use Myricom's clustering hardware, which roughly doubles the cost of hardware for each Linux system but is a key component of Cplant's fast internode communication.

Sandia uses Antarctica to simulate physical phenomena such as the effects of radioactive materials as well as for weather forecasting.