It's off season for supercomputers. We've come to expect two waves of buzz each year, coinciding with the semiannual release of the TOP500 list of supercomputer rankings, so the next rush of press releases should begin at the end of October and run until the updated list is formally issued on November 15th. I recently took advantage of the lull in hype to catch a beer with Donald Becker, co-founder of the original Beowulf project and Dr. Dobb's Journal's programmer of the year in 1999, and get his spin-free take on what the supercomputing industry still lacks.

Now, Don's got his own slant on things, of course. As the man who helped to invent the Linux cluster and who currently serves as CTO of Scyld Software, he believes strongly that commodity off-the-shelf hardware and open software standards will continue to dominate high performance computing. (Right now 304 systems in the TOP500 list are labeled as clusters, and nine of the top ten run Linux.) But he doesn't pretend that commodity machines can compete with the raw power of custom engineered supercomputers like IBM's Blue Gene/L, which is currently the most powerful computer in the world—and it's not even finished! "I always expected [commodity clusters] would handle the bulk of the high performance computing, but not all of it," Don says. "There's a real need for these custom designed, specialized hardware machines, but there's not enough market to sustain it. So the price/performance gets even worse."

After a few leading questions, I also got him to admit a bit of cynicism regarding the mass hand-waving that occurred in the U.S. when, two years ago, Japan's Earth Simulator clocked in at number one on the list. Until that time, American systems had always held the crown, and there was much (to my mind) jingoistic hysteria issued in response. "A good deal of that fear has been generated by the people who stand to gain from having the U.S. government continue to fund [them]," Don points out. "The Earth Simulator was from one standpoint a subsidy for the Japanese supercomputer market when the Japanese supercomputer market was struggling." U.S. interests wanted the same, and they got it; DARPA's $150 million High Productivity Computer Systems program was only one in a series of investments the government initiated in the wake of Earth Simulator. But the U.S. supercomputer market doesn't seem to be in any real danger; more than half of the high performance systems on the TOP500 list are installed here, and that percentage is on the rise.

Don's also lukewarm on the subject of grid computing. "For some organizations"—he gives high energy physics research groups as an example—"when the funding comes from a single source, and they need to share or access a single data set, it could be useful...it's primarily to access the data, not to access the resources." He's much more skeptical of the potential of massive grids: "They have to [take] machines that are separately administered over a wide area and get them to run an application without providing a mechanism to keep the environment consistent!"

Not that he thinks clusters are everybody's answer. "The biggest misconception is that you can just put a bunch of machines together and make a single threaded application run faster," he says. "You can make two apps run at the same time rather than sequentually." As a friend of mine says, it's the old 'how long does it take for TWO women to have a baby?' question. "Another misconception is that when you put together a cluster you can get shared memory very easily." Don points out that clusters have their own problems: higher overhead and cascaded failure. When one machine goes down, the whole application can crash; when one lock fails, the entire file system can be fubared.

And, as enthusiastic as Don is about Linux, he says more development is needed there too. Locking remains a big issue, as does process migration; Scyld believes in directive process migration, where it is the application's responsibility to request that it be moved within the cluster, but others design for transparent process migration and insulate the application from such activity. Still, he thinks Linux provides the best platform for these issues to be tested and hammered out.

"Linux gives the high performance computing world the opportunity to tune for the kinds of applications that run on high performance clusters," he says, pointing out that it's "a real challenge" to do that on a workstation that can't be modified. "There's a few different groups working on approaches to clusterwide process management...trying to agree well enough to get that as a standard part of the kernel. We have to agree that the hooks we put into the kernel will suit everyone fairly."

Scyld maintains a low-cost distro for groups looking to experiment with Linux clusters, but the current one is getting fairly old; a new one is in the works. Don also recommends that developers interested in Beowulf clusters check out the Beowulf.org mailing lists and BWBUG.org, the Baltimore-Washington Beowulf User Group. It's in these places, underneath the ebb and flow of TOP500 rankings and marketing hoopla, that real technical challenges are being solved.

Posted by Shannon Cochran at 04:53 PM | Permalink