Chuang Ren's goal: Maximum computing power to simulate problems of plasma physics

October 1, 2013

Video gamers know that a few hundred dollars invested in a graphics processing unit, or "graphics card," can make a world of difference – revving up the speed at which they can play the game on a computer and enhancing its visuals by several orders of magnitude.

Two Hajim School faculty members want to do the same thing for plasma physics, adapting existing simulation software so that it can be used with next generation graphic cards, multi-core processors and other advances in computer architecture.

Chuang Ren, associate professor of mechanical engineering, and Michael Huang, associate professor of electrical and computer engineering, say this would enable researchers and students alike to apply maximum computing power to the complexities of inertial confinement fusion, for example, or plasma astrophysics problems -- whether they're using supercomputers or desktops.

The goal of their project, funded with a $360,000, three-year grant from the National Science Foundation, is to allow "simulations never before possible on the world's most capable computers." They also hope to "greatly improve the kinds of calculations that can be done on an individual desktop machine. This will allow widespread use of sophisticated simulations by a large number of students and researchers."

"Both approaches," they believe, "will lead to new scientific insights."

There is an ongoing quest among plasma physicists to obtain more powerful computing capacity for their simulations. "The challenge we face is the huge scale disparity that exists between the physics we care about and the small scale physics we have to resolve," Ren said.

Consider, for example, inertial confinement fusion, where a simulation may involve a laser beam with a wavelength of 0.3 microns and a period of 3.3 femtoseconds; a target size of 1 millimeter and laser pulse duration of 1 nanosecond. (A micron is one millionth of a meter and a millimeter is one thousands of a meter; a femtosecond is one quadrillionth, or one millionth of one billionth, of a second and a nanosecond is one billionth of a second.)

To simulate that in 3D, "you have to have an enormous grid and run many, many steps," Ren said. "Even the largest computer in the world cannot simulate a laser shot in 3D from beginning to end. We are forced to simulate a patch of the plasma we look at, over a short period of time."

In the past, incremental improvements in central processing units (CPUs) provided some increased computational capacity for tackling these kinds of simulations, but further advances in CPUs "seem to be more and more difficult," Ren notes.

One alternative is to rely instead on GPUs, or graphic cards, which take over some of the most computation intensive portions of an application, while the remainder of the code still runs on the CPU.  GPUs "are probably now the most powerfully computing cores, because their development is driven by the gaming industry," Ren added. "They are packing more and more computing capability into them."

The plasma physics simulation software that Ren and Huang will be adapting for graphic cards is the so-called particle-in-cell (PIC) method. It's "a very versatile tool that has been used for different plasma physics problems," Ren explained. "Fortunately, 90 percent of the work is done by only 10 percent of the code. So we need to change only 10 percent to make it run on the GPUs."

Even so, "to make this work you have to have someone who knows the code and understands what you want to do, and someone who knows hardware," Ren added.

That is why the collaboration between Ren and Huang, an expert in various aspects of high-performance computer architecture, is so important.

They will also study how far the quest for greater computation accuracy should be carried before it produces diminishing returns. "You can always try to use finer meshes or more particles to be more accurate, but sometimes that's a waste of time," Ren said.  "You always need to balance accuracy with computation load. An important part of the study is to see when enough is enough."