Spring 2008
In This Issue:
Materials Modeling and Simulation
As our personal computers become more powerful, we expect even more from them – better graphics, more memory, faster performance. Scientific computing takes those demands a step further. When the fastest PC is still too slow, with too little memory to run complex calculations, Penn State researchers, and soon perhaps industry partners, can call on the Research Computing and Cyberinfrastructure (RCC) group at Penn State.
Located on the second floor of the Computer Building, the RCC group is a team of computer programmers and engineers who build and operate the research computer clusters for the University. These Linux based systems contain a total of about 1000 servers, or the equivalent of 1000 times the computing power of the best desktop models at the time they were put into service.
All computing is high performance, says Vijay Agarwala, director of the group. Today, anyone can buy a great deal of computing power for $500 to $2,000, he remarks. But in terms of research computing, he defines high performance as “something you can’t do easily on your desktop.” If you can speed up your number crunching by a factor of 10, he says, that can be considered high performance. He adds, “A hundred times would be better, and a thousand times would be great. But if we can’t get there, maybe we do have a combination of hardware and software the researcher doesn’t have.”
People who do theory usually benefit by doing analysis faster, Agarwala explains during my recent visit with his group. The speed of computation matters to theoreticians because they have more time to try out new theories, or to extend their theories into new areas. They can test a theory with a model instead of having to build an experiment. By model he means a set of mathematical equations that can be solved computationally. The model can then be tested against experimental results until the theorists trust the model and its predictive ability.
Research faculty, and some departments and institutes at Penn State have partnered with RCC to build the Lion clusters, from Lion-XL, the first cluster, to Lion-XJ, now being built. This gives those researchers priority access to a much higher level of computing than they would be able to afford on their own. Meanwhile, the RCC group builds and maintains the computing systems, schedules their use, and provides expert advice on the latest advances in high speed computing, leaving the researcher free to do research. The RCC is actively seeking new faculty partners as they strive to build a new cluster each year while pursuing new methods for increasing computational speed.
Crunching the Numbers
In recent years, parallel computing has overtaken serial computing as a less expensive way to gain more computing speed. Parallel computing links multiple processors to do computations simultaneously, as opposed to serial computation, which is run on a single central processing unit in sequential order. Improvements in high speed interconnects and multi-processor architecture have made parallel computing a better choice for many types of research, although Agarwala notes that it still requires about twenty times the number of processors to increase speed by ten times. This is a function of a consideration known as Amdahl’s Law, which shows that the speedup of parallel computing is limited by the portion of the computation that must be done serially.
“Our fastest machine can do 6.7 teraflops at its peak, and our aggregate capacity is 20 teraflops. A teraflop is a term that means one trillion floating point operations per second. A floating point operation might be 2+3=5,” he explains.
To compete with other top research universities, Penn State will need to increase its investment in computational infrastructure, Agarwala believes. In the next two years the U.S. will have a machine funded by the Dept. of Energy and the National Science Foundation that will have sustained computing capacity of one petaflop, which is 1000 teraflops. This is around 50 times faster than RCC’s current peak capacity. “Ask some materials scientists what they could do with 50 times more capacity,” Agarwala suggests.
Seeing and Sharing Data
The full functional name of the RCC group is High Performance Computing, Visualization, and Cyberinfrastructure. The visualization part refers to taking the numerical data collected by computation and turning it into a display. Turning data into an image can provide information that is easier to grasp, especially in an information dense field such as materials research.
“We can project very detailed images,” Agarwala says of their most advanced displays. “If we can see a picture, sometimes our minds will make a connection that we can’t see in the numbers.” The automotive and aerospace industry have been using computer modeling for many years to design and test new products without the expense of building physical models.
Modern science is far more collaborative than ever, and the ability to communicate over long distances and share data over the Internet has accelerated the collaborations. This is the cyberinfrastructure, which makes up the third element of the group’s functional name. Through the Access Grid Node at the Computer Building, a scalable environment for group-to-group voice and video teleconferencing, Penn State researchers can participate in meetings with colleagues worldwide.
In addition, the RCC group serves as the Penn State point of contact in the Coalition for Academic Scientific Computation, a 53-member organization representing many of the nation’s research universities and computing centers.
A Possible Benefit for Industry
Agarwala would like to make access to high speed computing available to industry in Pennsylvania, as part of the university’s historic Land Grant mission. Few small and medium size companies are likely to have high performance computer hardware, and even large companies with hardware may not be able to afford the kind of expertise represented by the RCC group. Agarwala is working with Rick Hoover, an industry expert in the Penn State Industrial Research Office with a strong background in materials and computer science, to judge the need among businesses for RCC’s excess capacity. “It will be our industrial outreach,” Agarwala says. “The university has a role to play in the economic viability of the state. If we can offer this service for a fee, then maybe we can grow our capacity by 10 percent and industry can use that extra capacity. But we will do this only after we have met the needs of the university.”
Contact:
To contact the Research Computing and Cyberinfrastructure group, visit them on the Web at http://gears.aset.psu.edu/.
Contact Rick Hoover at Validate to view contact info
