Riding Linux to the core

ComputerWorld Canada

Alessandro Forte wants to know the secrets of the planet, and he and his team are going to use an IBM Corp. Linux cluster to “journey” from the surface to the centre of the Earth to do it.

Forte, a professor of Earth Sciences at the University of Western Ontario (UWO) in London, Ont., said he wanted to put together as many processors as he could to solve difficult problems, which led him to IBM’s clusters. Plus, he noted, it’s a cheap alternative to buying an expensive mainframe.

The Earth Sciences department uses six node clusters – using an IBM eServer from the xSeries – with four processors in each node. Twenty-four 64-bit Itanium processors constitute the cluster. The processors work in parallel to address the researchers’ questions.

“The problem was the limitation in dynamic memory,” Forte said. “When you are doing a simulation of whole earth dynamics – from the surface to the centre – and you want to resolve the dynamics of the interior on a global scale with fine resolution, there’s no way you could ever do that on a single workstation.”

He added that the computation time would be impossibly long and the dynamic memory would not allow people to do those simulations. The 64-bit processors allow Forte and his team to pass twice as much information between the memory and the processor per clock cycle. They are using ultra-wide SCSI disks, with minimum speeds of 10,000 RPM and 70GB of capacity per disk.

Forte is also looking forward to 3D rendering, a technology that is advancing quickly on clusters running on Linux.

All of this is running on a Linux operating system, which Forte said was never in question. “Linux was always our choice because when you run simulations on a Linux operating system, they interface with processors very easily. With [other operating systems] you get a very inefficient result in terms of access to the processors.”

IBM has been a staunch supporter of Linux on its eServers for some time, said Dominic Lam, national eServer solution manager for high performance computing in Canada at IBM in Toronto. He said there is a trend to both Linux and clustering, especially in the scientific community. Forte noted that the applied math department at UWO is using clustering, as is the Astrophysics department at the University of Toronto.

Lam said IBM is gradually seeing large Canadian commercial customers implementing and asking about Linux clustering. He said people are seeing the performance benefits and that is sparking interest, less so than potential open source-style savings.

“It’s not a cost savings issue. There is no single Einstein in this world. That’s not to say that we’re less smarter nowadays, the research problems are just getting more complex…. It requires collaboration and the open source community, in general, fosters those collaborations,” he said.

Forte suggested that database managers would be looking more to Linux clustering as a “very robust” alternative with “very few bugs.” Linux machines can operate for weeks on end without needing to reboot, he said.

One other benefit, he added, is the failover for clustering. “With a cluster – which is just PCs connected together – depending on interconnect technology, which is quite flexible, if one element fails, the others continue to function.”

On the other hand with multi-processor mainframes, if one part broke down, the whole machine would be brought down to repair it. “This is really important for academic computing when you’re running simulations over days. You can’t afford to reboot,” Forte said.

Price was an issue for UWO, which found that the entire cluster, including peripherals, cables and installations, cost about $390,000, as opposed to millions for mainframe computers to do the same job, according to Forte.

Linux costs next to nothing, he said. “It’s a very democratizing influence.”

The project was helped by a grant from the Canadian Foundation for Innovation.