MIT’s Computational Reactor Physics Group (CRPG) supports leading-edge research across the Department of Nuclear Science and Engineering, providing new simulation and modeling capabilities for development and refinement of nuclear reactors. So why does the group’s co-leader, Benoit Forget, counsel his researchers to study the early days of computing for clues on how to accomplish their goals?
Forget’s answer speaks to CRPG’s critical role in helping advance new-generation high-performance exascale computing technology, which is essential for complex projects like detailed design and safety assessments of new classes of reactors. The technology, which promises a thousand-fold increase in power over today’s fastest computers, is being assessed by the Department of Energy’s Center for Exascale Simulation of Advanced Reactors (CESAR), a consortium of five national laboratories and three universities in which NSE Prof. Kord Smith, co-leader of CRPG, serves as chief scientist.
Unlike most supercomputer programs, in which hardware development has preceded software, CESAR (led by Argonne National Laboratory) is using a collaborative parallel approach that is expected to make exascale computing practical for complete reactor simulation by 2019. Teams led by Smith and Forget work closely with hardware designers as they develop code for machines that will use massive arrays of central processing units (CPUs) augmented by graphics processing units (GPUs).
“For the amount of processing power, GPUs are very cheap. One video card might equal tens to hundreds of desktops,” Forget explains. “The tradeoff is that there’s very little memory for each processor, so different programming strategies are required. It makes for a very interesting project because the architecture is so different.”
In the grand tradition of engineering, old solutions sometimes shed light on new problems. “In the early days of computing, programmers were limited in every aspect,” Forget says. “Now we’re limited by network communication, memory, pretty much everything except the processor. So we’re digging through old papers and code to understand the clever solutions they found, and extending that knowledge to overcome our own limitations and generate new ideas to make this work.”