|
Princeton
• Geowulf: Is this the future? |
|
|
|
|
|
|
|
|
On May 17 a delivery truck pulled up in front of Guyot Hall, and Hans-Peter Bunge's project to build a supercomputer was underway.
Bunge, assistant professor of geosciences, and several graduate students and undergraduates gathered around to help unload the cargo: 70 ordinary Pentium PCs. It was a Friday, and the group worked through the day hauling the computers to the third floor of Guyot where they arranged them in neat rows on gray metal shelving units. All through the weekend, late into the night, they ran cables and loaded software.
The result was, in many ways, a typical office computer system, minus all the monitors and keyboards -- ordinary PCs joined together by an off-the-shelf network device. But this system, called Geowulf, is among the 500 fastest scientific computers in the country, according to Bunge. And while others in its class would cost over $1 million, Geowulf's price tag, which included several workstations and auxiliary equipment, was $140,000.
Bunge hopes Geowulf will help him solve some fundamental problems in geosciences. But he also believes it will be a model for designing high-powered, inexpensive computers for use in fields from astrophysics to molecular biology.
"I've tried to take the most generic approach possible. Nothing risky, nothing fancy," says Bunge. The PCs themselves are "the most generic, vanilla-flavored."
Others at the University have been watching closely and plan to follow Bunge's lead. "I think this Geowulf structure is the future of scientific computing," said David Spergel, professor of astrophysics. "Hans-Peter has done some really neat things with his machine, and it's been great having him develop this because we've all been learning from it."
Astrophysical Sciences already has ordered 16 Pentium desktops to build its own supercomputer. The concept is attractive because standard office and home computers have become very powerful and very inexpensive at the same time. "We're basically just taking advantage of all the money that's going into the PC industry," said Spergel.
According to Bunge, the Geowulf project is in keeping with the spirit of the provost's recently completed initiative to upgrade the University's desktop computing capabilities and bring the desktops to bear on scientific computing problems. It also fits well with a new graduate teaching and research program, called PICASSO, aimed at integrating scientific computing across disciplines. "The ideas and the roots for all this were laid in a large part by Provost Ostriker," Bunge said. Another advantage of the Geowulf strategy is that when the supercomputer is eventually replaced, the individual machines could have continued life as desktop computers.
Movement of continents
The scientific problem Bunge is tackling is plate tectonics: What drives the movement of the continents, the creation of mountains? Plate tectonics is an unusual theory, Bunge says, because it is so well known, yet no one really understands how it works. Scientists know what happens, but not how or why it happens.
Early data taken by seismologists suggested that the earth's interior was too solid to move. Yet plate tectonics said that it must. So scientists took the contrary view that the rocks were acting like a fluid, and, like the atmosphere and the oceans, were moving around to transport -- or convect -- heat. As the interior cools, what appears to be solid and unmovable rock rises to the surface, loses its heat, then sinks again to the core. "You can look at the earth's interior as a body in motion," Bunge said.
Proving all this with conventional experiments on real materials is nearly impossible; the distances and time scales are too vast. That's where computers come in. Bunge uses a mathematical model that describes the earth's interior as a fluid that has very high viscosity and follows the laws of convection. Then it applies very long time scales.
Bunge began to run this model on conventional supercomputers. He divided the earth into 10 million cells, then instructed the computer to perform a series of 100 calculations having to do with mass, momentum and energy on each cell. These calculations get repeated thousands of times, which represents the passage of about a billion years. That adds up to trillions of calculations. On the computer screen, the result is an image of the earth with reds, greens and blues swirling up to the surface and back down, indicating the temperature on the surface or at any point inside. Running this simulation and adjusting for results that did not make sense has already yielded insights. It showed, for example, that rocks deep in the earth must be much stronger than was previously thought.
The success of Bunge's computer model prompted the American Museum of Natural History to incorporate it into its new Hall of the Planet Earth exhibit.
More like real Earth
Bunge's next step is to factor in characteristics of the earth's surface, so that the model will begin to look more like the real Earth. In particular, he needs calculations that describe the "breaking" of the earth's plates. Breaking is what allows plates to slide underneath each other at the fault boundaries. "That step is complicated, both conceptually and computationally," Bunge says.
Even with 10 million cells in his model, Bunge barely gets the level of detail he needs to draw conclusions; 100 million would be better. But just doubling the resolution means a 10-fold increase in calculations, he says. "So the name of the game is bigger computers."
That's where Geowulf comes in. At its current size, Geowulf allows Bunge to run his 10-million-cell simulations, which he used to have to do by renting time on $10 million-supercomputers. Expanding the system to handle bigger simulations is relatively easy; it's just a question of adding on more computers.
Bunge hopes to capitalize on Geowulf's power not just to advance his research, but to help change what he sees as a widespread misconception of earth sciences. "Earth science is really in a shift toward modeling and quantitative understanding of complex systems," Bunge says. "People don't think of geosciences as breaking new ground. People think: 'Earth science -- oh, that's where you look at rocks.' But earth science is undergoing a revolution."