Build your own supercomputer
Hewlett-Packard, together with a national laboratory in France, tried this recipe out. To the great surprise of many scientists, it worked. What they ended up with is the "I-Cluster," a Mandrake Linux-powered cluster of 225 simplified PCs from HP that has benchmarked its way into the list of the top 500 most powerful computers in the world.
At a technical session last summer, scientists from HP's own labs in Grenoble, France started talking to experts at the local INRIA Rhone-Alps (part of France's National Institute for Research in Computer Science) about the possibility of doing "something a little unusual": building a supercomputer out of standard hardware components such as might be found in a typical enterprise. They started with 100 of HP's e-PCs--simplified PCs with reduced expandability--and finally worked up to the present configuration of 225 nodes, which is near the cluster's physical limit.
HP and INRIA showed the system to journalists for the first time on Wednesday.
The version of the e-PC used for I-Cluster is sealed, meaning no hardware tweaks could be made to it. The experiment uses standard networking equipment as well. This means that, unlike with other clustered supercomputing projects, an organization such as a business could theoretically use the I-Cluster method to draw on idle computing power from around a company network to carry out computing-intensive tasks.
"These are really standard machines, we didn't even open the box," said Bruno Richard, program manager with HP Labs Grenoble.
Other clusters, like the United States' ASCI Red at Sandia National Laboratories, are comprised of heavily modified parts.
Sparking the cluster
There were formidable obstacles to face in getting the cluster running as if it were one device, Richard said. These included dealing with distributing functions like storage and network caching to general-purpose devices, and managing and programming the cluster.
"Our previous cluster was 12 machines," Richard said. "When you have 200 you have to rethink everything."
For example, even making simple software changes became a difficult task with so many machines to be altered. In the end, however, technicians devised tools capable of reinstalling every machine in the cluster from scratch in about 12 minutes, according to Richard.
The researchers plan to release the tools they developed as open-source software for anyone who might feel the urge to build a supercomputer. The whole project, minus network cabling, cost about 1.5m francs ($210,000).
The individual machines that made up the I-Cluster are now out of date, each running on 733MHz Pentium III processors with 256MB of RAM and a 15GB hard drive. HP introduced a faster version at the beginning of this month and will launch a Pentium 4 e-PC by the end of the year.
The e-PCs linked by fast Ethernet. Features like super-quiet cooling and low power consumption, originally designed for the corporate buyer, proved useful in the supercomputing environment too--the cluster runs surprisingly quietly and doesn't require anything more than standard air conditioning to keep it cool.
Judged using standard benchmarks, I-Cluster is ranked 385th worldwide and 15th in France for supercomputing. Richard said the experiment showed that there is a linear relationship between the number of nodes and performance, meaning that it's relatively simple to add or remove computing power depending on the task.
About 60 research teams worldwide are working on the system, with half running typical supercomputing tasks and the other half exploring how I-Cluster works.
Working to scale
The project shows that standard computing power--like the unused processing power on an office network--can be harnessed for serious computing work. In the business world, CAD designers and chemists are among those who need intensive computing power, Richard said. "You could gather the latent power from office PCs using this technique," he said. "We eventually want to scale it higher, to thousands of PCs."
Currently the hard limit for such a cluster is about 256 nodes, because of switching capacity, but that could be surpassed by linking several clusters that are physically near each other.
A more daunting task might be taking the model to a consumer environment, which, Richard pointed out, is full of often-dormant processors like those in printers and DVD players.
HP imagines "clouds" of devices, or "virtual entities", which discover and use the resources around a user. Richard said that supercomputing power could come in handy for certain tasks, like converting large video files from one format to another, that currently take a good amount of patience.
Other scientists predict that the practical difficulties of such a home network will prove difficult to solve. Brigitte Plateau, head of INRIA's APACHE parallel computing project, says that consumer need for such power probably wouldn't make it worth the effort that such a system would require.
"It is more likely that you would see an external service," she said.
HP's Richard said that the use of Linux--Mandrake 7.0 in this case--was important because low-level changes could be easily made to the software, and then the alterations could be shared freely with other scientists, something that would have required a special agreement with Microsoft if Windows had been used.
Plateau, whose APACHE project encompasses I-Cluster, said the lab is also working with Microsoft to port parallel computing applications to Windows. "We had to face heterogeneity, by spreading it over Linux and Windows too," she said. "It's not scientific, but technically it's good experience."