Clustered Linux: supercomputing on the cheap

When Jehangir Athwal and his colleagues wanted a system to do the heavy lifting at their bioscience software company, they were interested in a supercomputer system. But sticker price shock turned the IT professionals in a different direction--to Linux clustering.

The bill for the traditional supercomputer would have hit seven figures, while a Linux Beowulf cluster would run about $80,000.

That's the primary reason Athwal's company, LION Bioscience, is running a 44-node Beowulf cluster, powered by AMD Athlon Thunderbird processors, set up by Scyld Computing.

Athwal, a computational scientist at the German company's San Diego division, said a supercomputer offered more processing power, and its high-profile vendor promised a generous support plan. But those benefits still couldn't get the LION team beyond the price tag.

Besides, LION didn't need more processing power than the 44 nodes. If LION needs more parallel processing power to run bioscience software models and predictions, adding nodes is easy with a cluster setup; it's just a matter of a few more commodity PCs. With the supercomputer solution, the company would have to buy another supercomputer.

According to Linux vendors and industry analysts, cost-effectiveness is just one of several reasons prompting more enterprises to choose Linux systems.

Ease of scalability factor
The ease of scalability is another factor pushing Linux clustering beyond pure research institutions and into the business sector, especially in the "scalable performance cluster" category (as opposed to high-availability clustering and remote-access clustering).

In addition to companies like LION, petroleum exploration companies are embracing Linux clusters. Financial institutions are also using Linux clusters as small as eight nodes to run complex portfolio simulations and risk analysis, according to hardware and software vendors that include Dell and Sistina (a provider of file system and volume management software).

"Organizations having custom, computationally intensive applications are increasingly considering Linux configurations," explained Dan Kusnetzky, VP of system software for tech analyst firm IDC. "We're also seeing quite a bit of interest in low-cost Linux configurations for point-of-sale or point-of-solution systems that are replicated in many distributed sites."

Of course, Linux isn't the only player in the scalable performance clustering space, but Scyld CTO Donald Becker believes Linux makes good sense, especially when it comes to large clusters. Linux lets enterprises buy (or download) once and install everywhere, due to the open source aspect, instead of buying 2,000 licenses of Windows or proprietary Unix.

Becker, a cofounder of the Beowulf project at NASA in 1994, said Beowulf clusters have "much better price performance than traditional supercomputers" and they offer "as-needed" scalability. Advantages in addition to price include:

• Linux clusters run on commodity platforms.
• They offer long-term viability because there's no hardware or software vendor lock-in.
• They're tied to PC architecture, which has the fastest performance growth rate of any computing architecture.

Becker said Linux clustering initially scared potential users--they feared installing an operating system on dozens or hundreds of machines, and maintaining a huge number of PCs. But system management tools now available, featuring GUI or Web-based interfaces, let users "install once, execute everywhere," and monitor the whole system from one machine.

Kusnetzky noted that a lack of a "single system image" is still thwarting Linux in the clustering competition with Windows and Unix. His definition of "single system image" includes:

• Tools that let administrators see the whole multi-system configuration as a single domain.
• Tools that let developers create applications that work in a clustered environment without rewriting code.
• Tools that let nonparallel applications make some use of the multi-system configuration.
• Tools that let users queue batch and print jobs without having to know where the printing or computing happens.

Although Linux is catching up, Kusnetzky also explained that some needed administration, operations, and help desk tools, and some packaged applications, aren't yet available.

"IDC expects all of these issues to be addressed by suppliers and by the open-source community over the next couple of years," he said.

The analyst also noted that IBM, HP, and Dell have all announced sales of large, Linux-clustered configurations to compete with the supercomputing market. This will also help spur the needed tools currently missing.

While the price tag is likely the biggest factor organizations consider in choosing Linux clusters, Kusnetzky warned that the lack of tools is just as important.

"Linux is perceived to cost less and run on less costly hardware configurations. But organizations need to understand that the true cost of ownership includes costs of administration, operations, help desk, development, and other categories--staff-related costs factor in," he said, adding that those costs are why enterprises choose the more mature administration and operations software available on Unix and Windows.

While LION's Athwal is managing the new cluster on his own, he agrees somewhat with the analyst's statements. He recommends getting the best available cluster management tools instead of scrimping in that area.

"Definitely go with the bells and whistles," Athwal said. "If a node goes down, you don't want to spend all day looking for it."

Athwal also suggested that companies consider having a vendor handle the installation. "When we received the cluster from the vendor, I turned it on, and it started working."

Would you consider using Linux clusters for a supercomputer system? TalkBack below or e-mail us with your thoughts.