Beowulf: Supercomputers for the Masses
|"Beowulf is just starting. We haven't yet seen what we can get out of it."|
|"So far this year, Atipa has done 30% more Beowulf business than it did in all of 1999."|
Good for businessBeowulf may be in its infancy, but it is a baby that's getting a lot of attention. Even companies like Compaq Computer Corp., which has its own high-end clustering systems -- thanks to its acquisition of Digital Equipment Corp. -- aggressively markets Beowulf. While Compaq will gladly sell a customer TruCluster for Tru64 UNIX, Beowulf systems, particularly those based around the floating point-friendly Alpha chip, have been good business for Compaq. Some customers are using "literally thousands of Compaq Alpha servers with Beowulf," says Glenn Johnson, Director of Compaq's Linux Program Office.
Compaq's participation in the Beowulf market is as easy to understand as the appeal of the clusters to its customers. "Any industry that has a computational problem that can be parallelized is looking at Beowulf because you can assemble a hell of a lot of gigaflops for not a lot of money," says Johnson.
But Beowulf's price and open source advantages aren't always enough to win the business. To some extent, its open source development and distribution model have hindered its acceptance outside the academic and pure research realms. Even within those traditionally strong Beowulf markets, some major projects have opted for traditional supercomputers instead of Beowulf. Mark Seager, principal investigator for Lawrence Livermore Laboratory's ASCI White supercomputer says that Beowulf wasn't in the running for that system. ASCI White runs on RS/6000 servers, AIX operating system software, and Parallel Environment cluster code, all from IBM. At 8,192 processors, it is one of the world's largest supercomputers. Seager says that the "vendor solution" IBM could provide outweighed any possible advantages of a Beowulf-based supercomputer. "Beowulf is not a cohesive product," he says. "It's not what we would call production quality."
Companies such as Atipa Linux Solutions, Linux NetworX Inc., and VA Linux Systems Inc. are trying to remedy this, in part by providing end-to-end support for the Beowulf systems they sell, from hardware to OS to Beowulf itself. Still, even Beowulf founding developer Donald Becker says there is something to Seager's argument. Becker, who was part of the team that developed Beowulf at NASA beginning in 1993, says that when customers are looking for an off-the-shelf supercomputer, Beowulf is at a disadvantage.
The lack of integrated turnkey solutions has been "a real problem" for Beowulf, he says. Not every company interested in supercomputing has a resident expert. "In the academic and scientific community, there are a lot of people who can construct and maintain their own Beowulf systems, but people who do that are rather specialized, and you don't find that in the broader commercial world."
But some vendors are now seeing corporate customers who are taking a serious look at Beowulf for their "heavy hardware" needs. "We are seeing a big shiftit's the product lifecycle," says Brad Rutledge, public relations director for Linux NetworX. "We think Linux as well as clustering is going from the techie to the early-adopter stage," attracting companies that can afford to take technology bets in exchange for the price advantages and power of Beowulf systems.
And while no one can tell for sure, since the necessary software can be downloaded from the Internet for free, Beowulf clusters appear to be growing in number. Becker estimates that there are now thousands of Beowulf clusters, ranging from hobbyist 4 and 8-computer setups to monster systems with hundreds or thousands of nodes.
Three of these large systems are the "Los Lobos" Beowulf clusters, funded by the National Science Foundation. Like the ASCI White supercomputer, the NSF turned to IBM for hardware, but rather than paying top dollar for AIX and IBM Parallel Environment, Los Lobos will run Red Hat Linux and Beowulf. The first system, rolled out at the University of New Mexico, consists of 256 IBM Netfinity servers. Assuming all goes well, additional clusters will be deployed at Argonne National Laboratory, of Argonne, Ill., and the University of Illinois.
The Los Lobos systems will require additional software to accomplish their goals of modeling everything from quantum particle behavior to weather systems, but much of the new software will be released back into the Beowulf and Linux open source communities, including better I/O and filesystem interfaces for Beowulf. Project managers hope to make the three clusters available on demand to their research clientele, allowing a remote researcher to start a job on the cluster best suited to their needs.
Page 1 of 2