By Meredith W. Salisbury
If you happen to be driving in Ohio and come across one of those yellow Penske rental trucks, honk nicely: it might be Terry Lewis delivering yet another computer cluster.
Lewis, program coordinator at the Ohio Supercomputer Center, has been spending more time on the road in the last three years as OSC’s “Cluster Ohio” program takes hold. OSC, which provides production-scale compute resources via Internet connection to researchers across Ohio, came up with another way of spreading the wealth. Three years ago when they dismantled their first Linux cluster, a 128-node, 256-processor monster, to pave the way for an upgrade, it seemed like a waste to send the thing off to a Dumpster.
Just because the machines were no longer welcome at OSC didn’t mean they were obsolete: “It’s perfectly good gear,” Lewis says. “You could take one of these nodes … and use it for years and years. They’re that powerful.”
So OSC issued a statewide RFP, allowing researchers at any Ohio academic institution to compete for mini-clusters carved out of the original configuration. That process has continued and at press time a third RFP was open as OSC prepared to chop up its 400-plus-node cluster and bring in a new one with several hundred more nodes, Lewis says. In all, OSC has given out 22 of these hand-me-down clusters.
Technically speaking, life at OSC would be a whole lot easier if they didn’t have such an aversion to that Dumpster. In order to spawn the mini-supercomputers from Mama Cluster, engineers have to spend days or weeks, Lewis says, on building configurations that can work together — but independently of the rest of their siblings. “It takes a lot of extra effort,” Lewis says. Dealing with control, of course, is a major issue: how to spring several self-operating clusters from one mega-cluster that relied on a single front-end node and one power console? Teams that win each cluster typically have to agree to pony up a few thousand dollars —”we like to keep it below $10,000,” Lewis says — to pay for the additional routers and control equipment OSC has to buy to make the mini-cluster run independently.
When a cluster is delivered to a research team (literally by Lewis in a rented Penske truck), OSC engineers assemble the configuration and get it up and running. It’s connected to OSC headquarters so they can handle system administration, updates, and security patches, which means OSC has computing tentacles all over the state. Research teams are relatively free to swap their problems out to other sites if their nodes are saturated, and can remotely store data with their own security systems on OSC’s new 500-terabyte IBM storage array.
That spirit of connectivity will take a leap forward this fall when Ohio flips the switch on what it’s calling the Third Frontier Network — a statewide system that links compute resources using extra fiber-optic cable installed by telecoms in the late ’90s. “That’s significant for bioinformatics researchers because you can park data at OSC or you can use OSC’s resources to do things as if you had that giant cluster with hundreds or thousands of processors right next to you,” Lewis says.
Luis Actis at Miami University of Ohio says the cluster his school won helped pave the way for a foray into bioinformatics. The 32-processor machine lured the school’s first computational biology and proteomics hires and was key to the launch of its Center for the Advancement of Computational Research.
Michael Raymer, an assistant professor in biomedical sciences at Wright State University, currently has access to two clusters from Cluster Ohio. With the eight-node machine he won in 2001 — used for problems entailing full-genome scans for CPG islands or pattern recognition problems — a task that used to take up to 48 hours running overnight can now be completed in one or two hours. Raymer is particularly enthusiastic about being connected to other clusters around the state: “We are actually getting much more bang for our buck,” he says.