Clustering also seems tailored to Dell's corporate personality. Competitors such as IBM, Hewlett-Packard and Sun Microsystems spend millions annually in research and development honing their own microprocessors and operating systems for high-performance computing. Dell doesn't, instead relying almost exclusively on Intel, Red Hat and other companies for basic research. In clustering, though, that's not particularly a disadvantage. The Buffalo cluster can theoretically churn 5.6 trillion calculations per second (5.6 teraflops), making it one of the fastest clusters to date and conceivably putting it comfortably in the "Top 500" supercomputer list, Rooholamini said. Instead, the competitive differences come down to fine-tuning and assembly. Researchers need to figure out how many processors and how much memory should go into each server, and what sort of interconnects will provide optimal data flow. Dell not only participates in the tuning process, it also uses case experiences on other projects. "There are quite a few knobs that need to be adjusted," said Rooholamini. "Clusters also have a software component." "This is absolutely emerging as a core competency for them," said Jean Bozman, vice president of research at market analyst firm IDC. "What Dell has done is made it easy to acquire clusters. They are leveraging their business model to deliver scientific clusters more efficiently." Dell first began to promote Linux clustering about 1 1/2 years ago, Bozman said. While there are substantial differences between the huge university clusters and the smaller ones used in businesses, the company will increasingly begin to apply the lessons from one field to another. The proliferation of commercial databases into the cluster market will increase the size and heft of the commercial systems. SUNY Buffalo is the first university to receive Dell's research excellence award. The company has also funded ongoing research projects at the University of Texas, the Georgia Institute of Technology and the Pennsylvania State University, Rooholamini said. Choosing between a supercomputer and a cluster depends on the research project. Clusters work best when researchers analyse a large number of independent, divisible tasks. Fully analysing the structure of a single human protein within the human genetic code, for instance, could take 1,000 years on a single computer, Skolnick said. Using a cluster reduces the time to less than a year. By contrast, supercomputers work better on projects such as weather prediction and nuclear simulation, where the results of one calculation will affect the results of others. Although not as glamorous as supercomputers, clusters still have impressive power. The Buffalo cluster consists of 1,900 two-processor Pentium III servers, 100 two-processor Xeon servers, four servers for managing data traffic between the 2,000 computing nodes and a 14-terabyte storage area network from Dell and EMC and four servers for monitoring overall workloads. Switches from Cisco and Extreme Networks, meanwhile, link the system together. In all, it fills 41 computer racks and consumes about as much energy as it would take to heat 70 to 80 homes. "We're talking 80,000 pounds of computer here," said Skolnick.
See the Hardware News Section for the latest update on everything from MP3 players and PDAs to supercomputing.
Have your say instantly, and see what others have said. Go to the ZDNet news forum.
Let the editors know what you think in the Mailroom.
