As soon as the door of the freight elevator in Purdue’s math building opens, those exiting are assaulted by a blast of cold air and a pervasive hum.
It’s the sound of hundreds of cooling fans attached to bank upon bank of computers, all of which are churning away solving problems for the academics who buy space on the machines. Even in an era defined by miniaturization, these supercomputers are room-sized – the latter-day relatives of the first computers, like ENIAC.
But there is one thing that’s been miniaturized in 60 years – how long it takes to assemble such a machine, with its 700 processing nodes. Purdue's latest model was assembled in just a few hours.
Up on the dock of the math building, dozens of people unpack the computers from boxes and affix metal rails to each side, so they can be slid into the shelves in the basement.
Purdue Chief Information Officer Gerry McCartney says rather than each department buying its own machines, the school has moved much of the processing power to one place, and professors will use what they need.
“We have faculty that literally will queue up thousands of jobs and just say ‘I’m not in a desperate rush to get this done, but anytime in the next three or four months that there’s a free machine, I just want you to slap this job in there and just run it,” McCartney says.
Almost as soon as this year’s new machine was built, though, meteorology graduate student Kim Hoogewind began running weather models on it.
“I believe these runs here, we requested a fairly modest number of nodes – about six, I believe, at 20 processor apiece. So that’s 120 processors. So I believe this resolution, we got it done in less than an hour,” Hoogewind says.
On the computer monitors in her office are a weather prediction map on one screen and a many lines of code on the other. Hoogewind, like many grad students, first has to learn how to write the code necessary to make the supercomputer do the work. Only then can she have it output useful data. This, it turns out, is one of the big problems in supercomputing – writing the applications necessary to harness all that processing power. Professor Mike Baldwin, who oversees Hoogewind’s work, says most weather data generated today is basically thrown away.
“There’s probably 2-3 billion pieces of information that we get on a daily basis across the globe, mostly from satellites," he says. "And we’re actually only using a few percent of those data on a daily basis.”
This type of informational waste actually helps define how important building new machines is. As long as there are new problems to solve, supercomputers haven’t outstripped out their welcome – or the need to buy them every year.
"And what we aren’t event attempting to do today will be done in that room 20 years from now,” says Alan Chalker, who directs the Ohio Supercomputer Center at Ohio State University. He says weather is a good example of an ongoing research use for supercomputers.
“You still don’t know for sure ‘is it going to rain in your neighborhood this afternoon?” You get a chance of rain and if it’s 50-percent, it might or might not," Chalker says. "Over time, we’re going to be able to get more and more data, more and more precision, in those models, such that in 20 years, you’re going to know for a fact whether, in your neighborhood, it will rain and what time it’s going to rain.”
This computer cost Purdue about $4.6 million to build – about on par with what the school expects to spend annually for such a machine. The money comes, in part, from professors leasing the nodes from the school so they can use them whenever they want. When they’re not using their node, they can let others borrow it for a short while. Just having the computer there, says CIO Gerry McCartney, makes it easier to lure top talent to use it. But he still has to justify the cost.
“I will happily tell you we are a small fraction of the cost it would be to go outside. Now should that ever change, we will go outside. There’s no religion here," he says. "Then the imagination has to be: ‘now what can we do to help faculty do research and our students be more successful?’ Right now, that expresses itself as building these machines. In five years, it might be something completely different.”
Chalker says that’s a phenomenon that’s been borne out in the last 20 years.
“I can tell you that my smartphone and your smartphone are more powerful, by far, than the supercomputers we had installed in the mid-90s,” Chalker says.
So, much as it’s been since the days of the tube-based ENIAC, today’s supercomputer is tomorrow’s tech toy. But it’s a focus on what tomorrow’s supercomputers can do that defines the industry – and how schools like Purdue interact with the marketplace.