Two new NASA technologies have squeezed 10 times more power out of climate modelling supercomputers.

by Patrick L. Barry

For many people, faster computers mean better video games and quicker Internet surfing. But for decision makers grappling with the future of Earth's climate, faster computers have a very practical benefit: more realistic climate simulations that give more reliable predictions.

NASA scientists have managed to squeeze about 10 times more power out of cutting-edge "parallel" supercomputers through innovations in software and memory design. This leap in effective computing "muscle" - together with the data from NASA's Earth-observing satellites - enables greater realism and statistical confidence in simulations of global climate.

"That's something that we want to achieve, so that when policy makers have to make a decision involving hundreds of millions of dollars in the economy, they can say: "These simulations are something we have confidence in," says Dr. Ghassem Asrar, associate administrator for Earth science at NASA Headquarters in Washington, D.C.

Whether the question concerns the path of an approaching hurricane or the rise in global temperatures over the next century, predictions always carry some amount of uncertainty. But the computer "models" that produce the simulations can be improved so that this uncertainty is reduced.

Making these improvements will require all the computing power scientists can get their hands on.

To provide the immense "number crunching" power needed for demanding scientific applications such as climate simulation, some computer makers are turning to "parallel" designs with hundreds or thousands of processors in one supercomputer.

NASA  The bottom line shows how the increase in computing power (gigaflops) normally tapers off as the number of processors increases. The top line shows the performance on the same processors using the software tools developed by NASA.

The numbers on these machines make even the fastest desktop computers look like pocket calculators. For example, the newest supercomputer at NASA's Goddard Space Flight Centre boasts 512 processors each running at 400 MHz, 128 GB of RAM, 2,800 GB of disk space, and a peak performance of 409 gigaflops! (A "gigaflop" is a billion calculations per second.) A newer machine at NASA's Ames Research Centre will top even this with 1,024 processors.

But simply adding more processors doesn't guarantee a proportionate increase in effective power. In fact, the full potential of these parallel supercomputers still has not been tapped.

"So what's the problem? Each node (i.e. processor) has certain performance," Asrar explains. "Individually they perform well, but as you add them all together, as the number of nodes goes up, the overall efficiency degrades." For example, a system with 100 processors would not have 100 times the power of a single processor - the overall performance would be somewhat lower.

This loss of computing efficiency is a bit like what happens when people must work together to get a task done. Some effort must go into managing and coordinating the people involved - effort that's diverted away from producing anything - and even the productive workers must spend some amount of time communicating with each other. In a similar way, a supercomputer with more processors must use more of its power to coordinate those processors, and the increased communication between all the processors bogs the system down.

So the challenge was, how do you write the computer programs such that you get the maximum performance out of a single machine?" Asrar says.

Image courtesy NOAA Using faster computers, forecasters will be able to narrow the estimated path of hurricanes and perhaps save millions of dollars in unneeded evacuations.

For the past four years, scientists at NASA's Ames Research Centre have been working in partnership with computer maker Silicon Graphics, Inc., to tackle this problem. The fruits of their labour are two new technologies that increase the effective power of these machines by roughly an order of magnitude (that is, a factor of 10). Both technologies are freely available to the supercomputing community, are computer vendor independent, and are not specific to climate modeling.

The first of these technologies is a memory architecture called "single-image shared memory." In this design, all of the supercomputer's memory is used as one continuous memory space by all of the processors. (Other architectures distribute the memory among the processors.) This lets the processors exchange the messages needed to coordinate their efforts by accessing this "common ground" of memory. This scheme is more efficient than passing the messages directly between the processors, as most parallel supercomputers do.

But a new memory architecture needs software that knows how to make good use of it. The second innovation does just that. It is a software design tool called "multi-level parallelism." Software made using this tool can use the common pool of memory to break the problem being solved into both coarse-grained and fine-grained pieces, as needed, and compute these pieces in parallel. The single memory space gives more flexibility in dividing up the problem than other designs in which the memory is physically divided among the processors.

The extra computing power milked from the processors by these technologies will help NASA's Earth Science Enterprise make better models of Earth's climate.

These models work by dividing the atmosphere and oceans up into a 3-dimensional grid of boxes. These boxes are assigned values for temperature, moisture content, chemical content, and so on, and then the interactions between the boxes are calculated using equations from physics and chemistry. The result is an approximation of the real system.

With more computing power available, more of the physics of the real climate system can be incorporated into the models, and the atmosphere can be divided into more, smaller boxes. This makes the models more realistic, and the predictions they will produce will be of more interest on a regional scale.

Also, the ability to run these models faster will mean that more simulations can be performed, which will produce a larger pool of results. In statistical terms, this larger "population" will allow for a better analysis of the strength of the conclusions.

more The software tools developed by NASA and SGI can be used for other simulations, too. Shown here is a supercomputer model of a human protein.