Some of the IT world's top luminaries gathered at Los Alamos National Laboratory on Friday to witness the unveiling of a compact supercomputer that proponents say could provide the model for high-performance computing systems in the years ahead.
Bell, Torvalds usher in next wave of supercomputing
Gordon Bell, one of the original brains behind the minicomputer and Linus Torvalds, creator of the Linux operating system, joined a collection of scientists for the unveiling of the supercomputer — a Beowolf cluster called Green Destiny that was built from hundreds of so-called blade servers, compact servers stripped down to their most basic components.
Using server 'blades', Los Alamos scientists were able to build a system that is much smaller, consumes less power and is more cost-effective than typical supercomputers, according to Wu-chun Feng, team leader of the Radiant (Research and Development in Advanced Network Technology) group at Los Alamos.
"[Current supercomputers] are not addressing some of the fundamental issues that will be key for this coming decade," Feng said. "Simply doing bigger, faster machines is not good enough any more."
Different chips, different methods
Unlike many supercomputers which use specialised components and can fill entire buildings, the Green Destiny fits 240 server blades from RLX Technologies into a server rack that would fit inside most closets. The blades use low-powered processors from Transmeta, which made the famous Crusoe emulating processor, and a version of the Linux operating system.
The system consumes far less power than other supercomputers and should also require less maintenance, according to Feng. Using server blades meant trade-offs were made for overall CPU performance and internal bandwidth between components, but the research effort is still in its relatively early stages, he said.
A smaller version of Green Destiny called MetaBlade has been up and running for nine months without a failure, unlike other systems, which require "regular maintenance", Feng said.
Finding Crusoe no longer stranded
The engineers who developed the systems here said they preferred to use Crusoe processors from Transmeta because they don't rely on boosting transistor count to achieve faster performance, which they said it the case with chips from rivals AMD and Intel.
"In contrast to the traditional transistor-laden and hence power-hungry CPUs from AMD and Intel, the Crusoe CPU is fundamentally software-based with a small hardware core," Feng wrote in a white paper about his project, which is available in a PDF document on the web here.
"Because of the substantial difference in power dissipation, the Transmeta processor requires no active cooling, whereas a Pentium 4 [and most definitely an IA-64] processor can heat to the point of failure if its not aggressively cooled," he wrote.
Staying cool is only part of the solution
Some of the high-performance systems at Los Alamos use massive cooling systems to keep them working. The Green Destiny, however, sits in a dusty warehouse here where temperatures often exceed 80 degrees Fahrenheit.
Bell, who is now a senior researcher at Microsoft, attended the event because Green Destiny is in keeping with his passion for making high-powered computing systems as cost-effective and readily available as possible, he said.
"I was the guy who screwed up scientific computing by putting VAX out there," Bell said in a speech, referring to his work on DEC's VAX minicomputer, which in the 1970s helped lower the price bar for high-performance systems. "Groups of scientists could have their own computers out of the hands of big centres. This phenomenon is happening all over again."
Traditional measurement is not enough
In this vein, Feng also proposed that a new technique is needed for measuring the performance of supercomputers. Instead of looking primarily at how many calculations a system can run in a given amount of time, researchers should also consider factors such as downtime, size, price and maintenance requirements, he said.
Based on these measures, Feng said the smaller MetaBlade system would cost just over £22,500 to purchase and operate for four years, compared to more than £80,000 for a comparable Pentium 4-based system. The premium paid for the Pentium-based systems stems primarily from additional costs for system administration, the amount of power consumed and cooling costs, according to Feng.
Intel and AMD could come up with interesting techniques to cool their chips in the coming years, Feng said, but he urged fellow scientists to look hard at alternatives to speeding chips by adding more transistors, which tends to generate more heat. He didn't say that he had solved the problem with Green Destiny, but that his team's efforts are an important first move in the right direction.
"This type of project could serve as the foundation for the supercomputers of the future, but there is a lot of work to be done before then," he said.