Scale out with clustering

"It's funny with all the excitement about virtualization, people have sort of almost forgotten that clustering is a form of virtualization," Bozman says. "Clustering was one of the earliest forms of virtualization, in the sense that when an application is cluster-aware, it views all of the attached server nodes as being resources that it can use, as if it were on a big SMP [symmetric multiprocessing] machine.''

Users of large, high-powered Linux cluster systems say the mix of proprietary virtualization management software, along with low-cost hardware and free Linux, are opening up the processing-power floodgates.

CIS Hollywood is a digital special-effects house that produced digital images for "Pirates of the Caribbean," the fantasy epic "Eragon" and the most recent "X-Men" movie sequel, among dozens of other movies. Much of CIS Hollywood's rendering work -- in which large computer files are processed and crunched down into a viewable digital movie format -- is done on a cluster of 40 Linux PCs, running the free 64-bit version of the CentOS Linux distribution, which are managed by software from Linux Networx.

"The big key with Linux Networx is manageability," says Matt Ashton, systems manager for CIS Hollywood. "Instead of having to maintain individual nodes -- which can be done with a variety of scripts -- they've got all of that all set up to go. I can update all 40 machines with a few mouse clicks without having to do it by hand."

To CIS' users -- artists, graphic designers and computer technicians -- the Linux cluster appears as one large virtual machine. Fronting the cluster is a scheduling application written in-house, which distributes rendering jobs to the 40 machines. "Users don't interact with individual nodes," Ashton says. "They just submit jobs, and the queue management software takes care of it."

CIS has used a clustered, virtual rendering system for more than four years as a way to process the work of its artists more quickly and inexpensively. Ashton says nodes in the cluster -- dual-processor AMD Opteron boxes with 4GB of memory -- cost about US$4,000 each. CIS' large SMP Linux machines -- four-processor, dual-core machines with 32GB of memory -- cost between US$30,000 and US$40,000 each. The cost savings on a per-node basis is between US$2,000 and US$3,000 when scaling the system out, as opposed to up, he says.

PayPal, the online payment system owned by eBay, uses thousands of Linux machines to run its Web presence. The Web company replicates a single Linux/Apache image, bundled with its own transaction software, across these servers that appear as a single system to customers.

"Rather than have a monolithic box, we just have so many [nodes] that the breakages are irrelevant," says Matthew Mengerink, vice president of core technologies for PayPal.

However, few enterprises need the kind of computing power of a CIS Hollywood, or the scale of a global payment system, such as PayPal's.

Google is another example of the scale-out model, Steinman says. Its search engine runs on thousands of distributed Linux computers, which provide its signature fast, accurate search results. "But will an enterprise run its SAP platform on that model?" Steinman asks. "Probably not."