March 1999
Tapping the Source: LinuxWorld Expo news  Navigate    Home    Topical index    Archive  Print this story    Print-friendly version  Subscribe    Free e-mail alerts  Search          Contact Us    Masthead    Advertising Info    Writer Guidelines    Link to LinuxWorld  What do you think?    E-mail the editors

Newsstory

Linux servers at supercomputer speeds IBM engineers use Beowulf clustering to match Cray performance By Ed Scannell, InfoWorld

Using a subset of the Beowulf clustering technology, 17 of IBM's Netfinity servers containing 36 Pentium II chips and running an off-the-shelf copy of Linux matched the scalability and performance of a Cray supercomputer. The IBM system executed a computer graphics-rendering application called the PovRay benchmark.

The PovRay benchmark is intended to serve as a guide for the relative mathematical performance of a wide variety of chips, systems, and compilers. It is a ray-tracing image-rendering application by which a picture or image can be inserted in a movie such as Toy Story or Antz and subsequently be rendered with all shadows and rays of light appearing as they would fall relative to that picture or image.

"It is a big computational job. Ten years ago it would have taken a [Digital Equipment] VAX [minicomputer] 10 or 15 minutes to do. A Cray can do it in three seconds today,'' said Tom Figgatt, IBM's e-business segment manager, in Somers, N.Y.

During the demonstration, IBM's Linux-based supercomputer matched the current benchmark record of three seconds that was set by the Cray T3t-900-AC64, which had previously surpassed the second fastest time of 9 seconds.

The message IBM was trying to convey to users is that Linux has some innate capabilities for linking together parallel computers working in clusters -- not just working, but working robustly using existing hardware and software available off the shelf or on the Web.

"I think we showed how easily Linux clusters together and allows you to link multiple systems readily so you can spread your workload across multiple systems," Figgatt said.

In addition to the 17 servers, IBM used a 100MB Ethernet network and hub to connect the servers, and a piece of parallel computing software to ensure the system's computations all connected. As for the copy of Red Hat's Linux, IBM purchased it at a local Barnes & Noble bookstore the day before the demonstration

"We got the Linux by walking to the local store where we bought a copy of the book Linux Unleashed for $39.95, and in the back of the book was a copy of Red Hat Linux," Figgatt said. "Part of the message here is we just took Linux off the shelf -- well, Linux out of the book -- and were able to do some very interesting clustering with it."

Although the application demonstrated would be considered exotic by most Fortune 1000 companies, IBM believes many commercial accounts need this level of computing power for many of its existing and upcoming Internet-commerce applications.

The advantage of the IBM-based system over the Cray, of course, is its more attractive price performance, company officials said. The Netfinity/Linux benchmark was done on approximately $150,000 worth of equipment; the cost of the Cray used was $5.5 million, they said.

IBM also used the demonstration to flex the muscles of its X-architecture features and capabilities, which now are included in all of the company's servers up to the mainframe-class machines. For example, during one of the rendering demonstrations IBM took one of the servers offline. The screen performing the rendering missed several pixels during the fail-over but had filled them in by the time the rendering was complete.

The benchmark results are now available at www.haveland.com/povbench. Users need to click on the button that is labeled "list all parallel results."

Discuss this article in the LinuxWorld forums (4  postings)
(Read our forums FAQ to learn more.)