Sandia LabNews

Portable cluster computing

Sandians make cluster computing portable

Even cluster computers may become smaller and more portable under the inspiration of a demonstration system put together by folks in Sandia’s Embedded Reasoning Institute.

A cluster computer normally combines several desktop-sized personal computers to work together on large problems in a fast, affordable way, operating on such software systems as the widely used Linux.

"A lot of people thought it was really cool," says Rob Armstrong (8915), who showed their Linux cluster of four CPUs, just 13 inches in its largest dimension, at Supercomputing 2001 in November. (Overall measurements are 5.3″ x 5.3″ x 13″.) Rob just stowed the commodity cluster in the overhead airplane bin on his flight to the meeting in Denver and back.

The next month, Mitch Williams (8945), who oversaw its creation, won a "Work in Progress" presentation at the annual Large Installation System Administration conference.

"It’s being touted as a small Linux cluster with portability for tutorials, demonstrations, and road shows," he says. "The alternative for folks who want to go to an exhibition with a cluster computer is to spend months coordinating the shipping of a rack of perhaps eight full-size personal computers.

"It’s no fun," says Rob, who has done this. "It’s very cool that you can get it into that small space." Both Oak Ridge and Los Alamos national laboratories are now building their own versions, inspired by this early example.

The cluster uses the PC104 "embedded system" hardware standard, with units stacked like a club sandwich. All of the components, except the see-through Lexan case, were purchased off-the-shelf from embedded system vendors. Although the four CPUs don’t offer the power and speed of a larger cluster, they will still run software created for cluster computing, so specialists can demonstrate their code at a remote location much like other presenters use laptop computers on the road to present their wares.

Besides the portability advantage, Rob says, the approach is also attractive to vendors seeking a more compact way to offer commodity clusters. "It’s certainly far more dense than hooking PCs with a hub," he says.

The system was integrated for Sandia by Parvus Corp. in Salt Lake City in October, which Mitch visited so he could test the unit on-site. The CPUs were provided by Advanced Digital Logic.

Mitch created a web site for those interested in creating their own "cluster in a breadbox":

Next to larger research clusters like Sandia’s CPlant, he says, "it’s just a toy — but it’s a very interesting toy to the people who are programming CPlant."