Dr. Dobb's TechNetcast shows


h o m e

s c h e d u l e

a r c h i v e s

n e w s

c h a t

f a q

t o o l s

a b o u t


dr. dobbs journal


technetcast 1998-06-12

NASA Disables Beowulf Project

Missiles to China and
Linux Parallel Machines


Click here for video interview with Don Becker

Don Becker left for vacation a few weeks ago with peace of mind. Beowulf, the Linux parallel computer project he spearheads was gaining momentum. More importantly, the technology was stable and performed beyond all expectations. An increasing number of sites were setting up and testing Beowulf clusters. An arrangement with Red Hat Software provided exposure and a valuable distribution channel . This cutting-edge technology would now be more easily available to Linux users everywhere. The mainstream computer press was also paying attention: PC Week Labs tested Extreme Linux (the Red Hat distribution) and found it to provide "incredible results" (the piece also criticized the difficulty of setting up and configuring a working cluster). It was a good time to take a break.

He returned from his vacation to find that the project site had been shut down and future of the project was in doubt. And the perpetrators belonged to his own organization, NASA.

Don works at CESDIS (Center of Excellence in Space Data and Information Sciences), a NASA contractor, at the Goddard Space Flight Center in Maryland where he started Beowulf in 1994. (For a history of the project, and a complete review of the technology, see the Extreme Linux site). The project aims to provide the technology and resources necessary to architect, set up and run high-performance Linux clusters. This is achieved by both using available technologies -such as existing parallel processing and message passing environments (such as PVM and MPI), and by developing home grown solutions where needed. In particular, Don worked on enhancements to the kernel network subsystem to support faster I/O on high-speed networks. He wrote Ethernet device drivers. A distributed shared memory package was also developed. In addition to the technology, Beowulf also provides a hub for developers, administrators and testers to obtain and exchange information about their activities.

The results can be spectacular. There are currently a few dozen Beowulf clusters running at various sites around the world. The Hive, the cluster running at Goddard, is a 64 dual processor node computer and achieves flop rates in excess of 10 Gflops (this rate are standard across most Beowulf implementations). 100+ node clusters have been successfully set up and operated. 16 node clusters are common.

Don is in part the victim of bad timing. An unnamed source close to the project attributes the action to "those very bureaucrats at NASA who screwed up in providing spare missile parts to China". Apparently, the heat the agency received for its involvement in the current Loral-technology transfer to China affair caused it to be extra-sensitive to U.S. export regulations. The exposure received by the Beowulf put the project on the radar screens of the compliance officers. According to the same source, the intensity and speed with which the agency moved to paralyze the project is in direct proportion with the flak it received in the Loral case.

The problem is that Beowulf clusters can achieve processing power well beyond the 7,000 MTOPS (Million Theoretical Operations per Second) ceiling imposed by the U.S. supercomputer export laws for certain countries. If congress has its way, this limit will be brought down to 2,000 as part of the defense package currently being evaluated. A typical Pentium 200 MHz processor roughly achieves 200 MTOPS. A 300-MHz Pentium II chip operates at 1000 MTOPS. Assuming linear scalability, a 35 node, 200 MhZ cluster of would run afoul export laws.

Another internal concern at NASA is that Beowulf may be competing with technologies offered by vendors who currently do business with the agency.

The Beowulf site is back up, but a message indicates that "some of the software on this site has been rendered unavailable pending export control review". In the meantime, as any well-designed distributed system, Beowulf lives on. A mirror site has been set up. As is more and more the case, slashdot is a focal point for the Linux community and serves timely news and commentary about the affair. The technology is also available from the Red Hat distribution. R&D sites around the country continue to operate Beowulf machines.

As for Don, he is allegedly committed to pursuing his research at the Goddard Space Center or elsewhere. Other NASA centers are currently test sites for the project and may provide a haven until the project fades again from the memory of the bureaucrats .

--PL