IBM builds Linux supercluster
by Marc Songini
ALBUQUERQUE, N.M. – A system of computers from IBM may soon let users run Linux applications at supercomputer speeds.
Last week, IBM announced it had jointly created a cluster of 256 Netfinity two-way servers running Red Hat Linux at the University of New Mexico. The resulting prototype system, called “Los Lobos,” can handle 375 billion operations per second - making it the 24th fastest supercomputer in the world, IBM says. The fastest machine in that class is reputed to be the ASCI Red SuperComputer, based in Sandia National Laboratories, which handles one trillion operations per second.
The $1.5 million Los Lobos system, expected to go into production this summer, will run complex math, physics, chemistry and genetics applications for scientists and researchers located all over the country who will be able to tap Los Lobos via the Internet. IBM is considering rolling out a commercial version of Los Lobos software in the future, says Tom Bradicich, an IBM executive.
Experts say that while Linux is popular with certain technical users and is widely used as a basis for Web sites, it is not considered to have the muscle for handling industrial-strength business tasks. Los Lobos may change that perception.
Los Lobos isn’t the only example of large Linux clusters. Alta Technology claims its parallel servers will let users tie together more than a 1,000 Intel processors at a time to run Linux applications.
IBM already offers a supercomputer, the RS/6000 SP, that doesn’t run Linux, but rather IBM’s AIX. Nevertheless, a Linux supercomputer would be a lot less expensive than an SP, which can cost millions of dollars.
Teams of programmers from IBM and the university wrote the software to install, configure and manage the Los Lobos cluster, using open source Beowulf code and tools from third parties such as the Extreme Linux community.
Within the Linux supercluster, one dedicated server acts as the master scheduler, deciding which individual machines will be assigned which tasks and when. On the network side, the lab is using Myrinet, a proprietary network technology from Myricom, to link processors and server chassis.
Similar Linux clusters are going to become commonplace for shops needing a lot of computing power, according to Frank Gilfeather, the University of New Mexico’s director of high-performance computing. “This is the future,” he says.