To assess the stability and safety of proposed nuclear reactor designs, UC Berkeley nuclear engineers utilize the campus High Performance Computing (HPC) cluster, Savio, to predictively model the pathways of neutrons as they collide with atoms in the nuclear fuel. Kelly Rowland, a PhD student in Professor Rachel Slaybaugh’s lab, and a domain consultant with Berkeley Research Computing (BRC), develops and tests computational methods for simulating neutron motion. Her research is supported by BRC’s HPC services, allowing her to run complex calculations associated with the neutron transport equation in parallel, building accurate models of nuclear systems on the aggregation of individual neutron pathways.
BRC’s Steve Masover and Olivia Habern spoke with Rowland about her research; her role as a consultant; and the applications of BRC services for her individual projects and the Department of Nuclear Engineering.
Computational Methods of Neutron Transport Simulation
Nuclear reactors produce energy through a chain of contained nuclear fission events, in which the nucleus of an atom of fuel absorbs a neutron and explodes into more neutrons and fission fragments. Safe, sustainable nuclear reactors balance net neutrons gained and lost between fission events. This critical state can be described by a numerical quantity, the k value, derived from solutions to the neutron transport equation.
Rowland utilizes a predictive model of neutron transport to reveal what would happen at the atomic level under a range of possible reactor conditions. Monte Carlo simulations model the pathways of millions of individual neutrons and the results of their modeled collisions. These results are aggregated to determine the k value of the overall system. A k value of 1 represents an existing equilibrium between net neutrons gained and lost in fission events, and suggests that the nuclear reaction will neither fizzle out and die, nor accelerate to the point of excess energy production, i.e., explosion.
This numerical assessment of reactor designs via neutron transport simulation can only be accomplished efficiently on a high performance computing cluster. Rowland puts it simply: “I need to get these k values right, and I need to get them right quickly.” She explains that Monte Carlo simulations assume that neutron histories are independent of each other, so the calculations for many neutron paths can be run in parallel. HPC clusters such as Savio are designed to optimize this kind of parallel computing, in which large problems (like solving for the k value of the overall nuclear system) are broken down into small problems (like the outcome of each individual neutron pathway) and calculated simultaneously. The general purpose Graphic Processing Units (GPUs) on Savio can be leveraged to accelerate these Monte Carlo neutron transport calculations by harnessing the power of thousands of GPU cores to further parallelize Rowland’s calculations.
Savio is also uniquely helpful with benchmarking Rowland’s code, comparing it against other production-level code used for reactor physics calculations. Benchmarking tests the accuracy and speed of her code in doing criticality calculations. The code Rowland works with is only for use with GPUs; however, the production-level code she benchmarks against is only compatible with CPUs. Because Savio includes both classes of nodes, she can benchmark the code smoothly, making fair speed comparisons on GPU and CPU hardware of similar generations.
Dissertation Work and Future Computing Needs
Rowland’s dissertation work will focus on developing and testing hybrid methods of neutron transport simulation, combining and comparing Monte Carlo and deterministic solutions. While Monte Carlo models the trajectories of individual neutrons at exact coordinates in the reactor, deterministic methods approximate neutron behavior in a finite number of small cells that cumulatively represent the reactor system. While Monte Carlo methods are accurate, deterministic methods are fast. Rowland wants to create a method that unites the best qualities of each class of simulation.
Her project will continue to involve BRC’s HPC services, coupled with a possible application of the emerging software technology Singularity, which provides a way to package artifacts of computation workflows -- one or more Linux applications, with all their dependencies -- and run them successfully in a variety of other environments. Singularity will allow Rowland and the coders she is collaborating with to use bleeding-edge versions of the software they need, and still run their simulations on Savio, which is provisioned with older, more proven software modules. As Rowland explains, “new versions of software can be a bit unstable, so it makes sense for system administrators of an HPC cluster to be wary of having software packages on their machines that are always brand new. Singularity comes in here. My group can build the software stack we want off of Savio, giving ourselves a computing environment configured for our specific needs, but then bring it to Savio and make use of the cluster’s compute capability.”
BRC Consulting
As a researcher and a domain consultant embedded in the department of Nuclear Engineering, Rowland is in the unique position of being both a user and a provider of BRC services. She has played a vital role in the process of porting users of her department’s cluster over to Savio, a transition that has yielded great results for the researchers she supports. When asked about her experience consulting, Rowland remarked on the disparity between undergraduate training and graduate expectation with regard to scientific computing: “I support a broad range of research within the department of Nuclear Engineering, but I also support scholars with a broad range of scientific computing experience. Some researchers are self-sufficient in regards to computing, and others don’t know what a terminal or a command line are. This is really hard because when you enter graduate school everyone suddenly expects you to know how to use a computer, and how to use it well, but no one ever stops to teach it to you. I had to teach myself computing. It was like grasping at straws on the Internet. With BRC consulting we are attempting to bridge the clear gap between undergraduate training and graduate expectation for computational research. By helping to install software packages or running software architecture workshops, I’m trying to be there for other graduate students. I want to be visible to them as a resource that can help them do their research.”
If you are interested in learning more about our campus HPC cluster, Savio, or about BRC consulting, please contact us at research-it@berkeley.edu.