This article can also be found in the Premium Editorial Download "Storage magazine: Storage products of the year 2003."
Download it now to read this article plus other related content.
The lure of cheap Linux clusters is getting stronger. According to Steve Feldman, director of software development at CD Adapco, which makes software to calculate computational fluid dynamics (CFD), "every day we get a call from someone about Linux clusters."
That's because a Linux cluster made up of inexpensive one- or two-way Intel blades costs a fraction of a single equivalent large RISC-based symmetric multiprocessing (SMP) server.
But without a clustered file system to provide a single data store to each of the parallel blades, management of the cluster is a real headache, Feldman says. To that end, CD Adapco has deployed its software on a Fujitsu's hpcLine of Intel-based servers running Sistina Software's Global File System (GFS) a clustered file system.
Without a clustered file system providing centralized disk management, each CFD "run" requires that the full dataset be copied to a local disk associated with each processor in the cluster. Then after the run is completed, results are merged and copied to a central location. By implementing GFS, Feldman says, CD Adapco was able to eliminate this data movement step.
In the days before GFS, CD Adapco had scripted these steps, which minimized the amount of manual labor an administrator had to perform. And it all worked fine, "until something went wrong." For example, an entire run might fail if a processor's disk was full. Runs take anywhere between twenty minutes and a month, Feldman says.
Of course, injecting a clustered file system like GFS into your Linux cluster does add some cost--but only about 10%, estimates Feldman. "We don't do elaborate cost analysis," he says. "But it's not doubling the cost, or we wouldn't do it." One thing is for sure, though: "it makes it a better system."
This was first published in January 2004