Climate research lab picks SGI for forecasting duty

The National Oceanic and Atmospheric Administration has agreed to install 10 Origin 3800 computers that will be used for forecasting heavy-duty weather phenomena such as hurricanes and climatic trends such as global warming.

Raytheon, long a supplier for big-ticket government contracts, won the four-year, $34 million contract to supply NOAA with the SGI computers, NOAA said. SGI will supply eight 128-processor Origin 3800 machines and two 64-processor machines, SGI said. The systems will be installed at NOAA's Geophysical Fluid Dynamics Laboratory in Princeton, N.J. After four years, the contract may be extended with another $33 million purchase.

The computers will be used to conduct a number of computationally intense simulations, including modeling hurricanes and El Niño weather patterns, said Brian Gross, a GFDL physical scientist who helped select the new computers.

"We build and execute numerical models of the atmosphere and ocean," he said.

The Origin 3000 series from SGI, introduced in July, is the latest hardware design the Mountain View, Calif., company has launched in a bid to return to profitability. With the new design, SGI has almost completely left behind its once-grand ambitions to sell general-purpose servers, choosing instead to focus on its core, niche market of technical customers.

The SGI machines will take over for a number of Cray computers, he said. Compared to the Cray computers, the new systems will perform simulations four times as fast, he added, a key advantage for the computationally intensive task of simulating a world sliced up into a large number of interacting regions.

In March, SGI sold Cray to Tera Computer, which adopted the Cray name.

SGI is not the only one angling for the climate-modeling market. IBM is making inroads with its RS/6000 line, now renamed eServer P series, which is used at another federally funded lab, the National Center for Atmospheric Research.

In addition, the Navy in August bought an $18 million IBM machine to model oceanic weather. And in 1998 IBM won a $35.6 million bid at the National Weather Service.

Meanwhile, Compaq Computer is making a push of its own into technical computing, winning a bid at the Los Alamos National Laboratory to build what is expected to become the world's fastest supercomputer in a contract worth $200 million. The machine, called "Q," will be used to simulate the physics of nuclear explosions, a task surprisingly similar to modeling the weather.

Modeling of nuclear explosions or weather patterns typically begins by dividing a region into a three-dimensional grid, creating a stack of cells. Each cell has its own physical characteristics, such as temperature, pressure and humidity, and interacts with its neighbors according to physical laws.

For example, high pressure in one cell may spread to the next, leading to a shock wave.

The more cells that are used, the more accurately the model portrays reality, but also the heavier the computational load.

Software for mathematical simulations therefore is tuned to take advantage of the unique designs of high-end computers. But that careful adjustment makes it difficult to move software from one computer to another--from, say, a Cray to an SGI machine.

The NOAA lab does not have to worry about the translation difficulty, however, because it is moving to a new software model, Gross said. The new software has been written to be more flexible, with a framework that can accommodate several different components to handle atmospheric or oceanic models.

"It's easier to switch in and out separate pieces of climate and weather models, such as a land or sea ice model," he said.

One advantage of the system is that it is better at handling the differences between modeling the ocean and the atmosphere, he said. Typically, ocean models divide the sea into smaller volumes that react very slowly. The atmosphere, by contrast, uses much larger cells but shorter time scales.

It can be awkward cramming these two different approaches into a single computer model, Gross said. But the modular approach "to a large extent lets us not worry about how the atmosphere and the ocean talk to each other within the computer program," he said.