Corporate Applications for SupercomputingBy Kim S. Nash | Posted 2002-12-04 Email Print
CGG Americas has put in a PC-cluster where it once had supercomputers as it seeks new spots to drill. There's no turning back.
Corporate Applications for Supercomputing
With supercomputers, a typical seismic imaging job might take four months because of unexpected failures. Now it's down to three weeks, says Derick Deaton, executive vice president of CGG Americas. "Companies are deciding where to drill in the years to come and 'months' doesn't sound like much," Deaton says. "But it can cost millions" in foregone revenue.
Another benefit: The tools to manage a cluster are less expensive and require less-specialized skills. Where it used to take five or six technologists to maintain CGG Americas' supercomputers, the cluster demands just three. Plus, it's easier to find talent, says Laurent Clerc, the technology architect who manages the cluster at CGG Americas. The company's massive Linux cluster is a rarity in the corporate realm. "People find it fascinating to put their hands on it," he says. "The main thing is to tell them this is not a super big toy for them."
Though most supercomputing has been done at universities and government laboratories, some corporate applications are getting complicated enough for such computing strength. Consider, for example, the data gyrations necessary to predict customer behavior accurately. Amazon.com last year installed a supercomputer-class machine from Hewlett-Packard, called a SuperDome. State Farm Insurance runs customer databases on two IBM SP Power3 supercomputers.
But in the corporate realm, high-performance clusters, when used at all, are used mainly in scientific and research work. Ford Motor Co.'s research unit has used a 256-processor cluster of Sun Microsystems machines for two years. Honda Research and Development created its own cluster out of Intel parts early this year. Drug researchers Syrrx and Inpharmatica both installed clusters of IBM Netfinity servers this year. In the oil industry, the Royal Dutch/Shell Group and two CGG rivalsWesternGeco in the U.K. and Petroleum Geo-Services in Norwayuse Netfinity clusters.
"The power is there, in a cluster. There is no doubt about that now," says International Data Corp. analyst Jean Bozman.
CGG Americas has spent more than $7 million on its Dell cluster, Deaton says. That includes the PCs, racks, accessories and the Linux licenses from Red Hat Inc., in Raleigh, N.C. All told, the total cost of ownership of a cluster is half that of the old supercomputer platform, says Clerc, the technology architect.
Clusters, however, aren't trouble-free. PC hardware is constantly improving, making machines obsolete that are only a few months old. CGG Americas has had to budget for the nuisance and cost of steady upgrades, even while continuously enlarging the cluster. Building efficient clusters from thousands of PC servers and their attendant accessories takes a lot of time and fanatical attention to detail.
Clerc takes pride in how he designed his cluster. A single rack, about seven feet tall and 18 inches wide, holds 32 dual-processor servers, one switch to control data moving between the processors and a so-called "concentrator" that also helps manage network flow. A master rack contains spare systems and various software tools, including some from Dell and some proprietary to CGG, to monitor and maintain the cluster and schedule jobs.
"There are so many pieces in a cluster that if you start neglecting one or the other, it will break down," Clerc says. "You have to be a maniac about how you build it."
Click here to see a side-by-side comparison of two methods of crunching numberssymmetric multiprocessing and parallel processing.