DOE Launches Storage Effort

The U.S. Department of Energy (DOE) is investing $11 million in a project to solve the storage problems of high-speed supercomputers, a move which may ultimately help enterprises handle their own spiraling storage demands.

The money is being used to fund the Petascale Data Storage Institute a research group of universities and government labs, led by Carnegie Mellon University, which will tackle the storage challenges posed by next generation supercomputers. With users now planning to deploy "petaflop" machines capable of a quadrillion (a million billion) calculations a second, the strain on storage will be immense, warns Garth Gibson, associate professor of computer science at Carnegie Mellon, and head of the Institute.

"The faster the processor, the faster you can absorb and create data," explains Gibson, who is also the CTO of clustering specialist Panasas, adding that a petaflop supercomputer may require hundreds of thousands of magnetic disks for storage. Such a large number of components, according to Gibson, increases the likelihood of a system failure.

Researchers at Los Alamos National Lab, home to one of the largest supercomputers in the world, have already warned that the machines typically suffer failures once or twice a day. Once the technology reaches the petaflop level, warns Los Alamos, this failure rate could jump to once every few minutes.

Supercomputers, although traditionally the preserve of universities and shadowy government research labs, are increasingly making their presence felt in the enterprise.Although research institutes dominate the high-end of the Top 500 list of the world's largest supercomputers, well over 100 enterprises also featured. These include Paramount Pictures , HSBC, BMW, and a slew of unnamed manufacturers, gaming companies, and digital media firms. (See High-Speed Links Favor Ethernet.)

"Any place where you're getting Linux clusters, you're going to see scalable, distributed, storage systems," says Gibson. "It could be in seismic analysis, image rendering for movies, airflow design for airplanes and cars, designing chips, and financial modeling on Wall Street."

The Data Storage Institute will focus on a number of key areas, such as collecting supercomputer failure data, and developing best practices and standards, which may eventually be adopted by the commercial sector. Gibson explained, for example, that the Institute is already working with IBM, NetApp, Sun, EMC, his own firm, and the University of Michigan to develop a prototype for Parallel Network File System (PNFS), a high bandwidth version of the NFS protocol specifically for supercomputing.

Unlike the current version of the protocol, PNFS will enable a client device to pull data from 1,000 IP addresses, as opposed to just one, according to Gibson. This additional capability should reduce the potential for network bottlenecks.

The Institute is also developing what are described as "self-star" systems, which are software-based techniques for automatically managing the relationship between supercomputers and storage. "It will cover areas such as data mining, data migration, and power and temperature monitoring," explains Gibson.There is already a precedent for this type of initiative. According to Gibson, the open source Lustre file system was born out of a similar project funded by the Defense Advanced Research Projects Agency (DARPA) in the mid-nineties. (See NAS Roadmap and CFS Updates File System.)

The Data Storage Institute brings together computer scientists at Carnegie Mellon, the University of California at Santa Cruz, and the University of Michigan. The initiative also includes the DOE's Los Alamos, Sandia, Oak Ridge, Lawrence Livermore, and Pacific Northwest national labs.

James Rogers, Senior Editor, Byte and Switch