SUBJECT: Ph.D. Dissertation Defense
BY: Paul Burke
TIME: Thursday, April 22, 2021, 10:30 a.m.
PLACE:, Online
TITLE: COMET-GPU: A GPGPU-Enabled Deterministic Solver for the Continuous-Energy Coarse Mesh Transport Method (COMET)
COMMITTEE: Dr. Farzad Rahnema, Co-Chair (NRE)
Dr. Umit Catalyurek, Co-Chair (CSE)
Dr. Bojan Petrovic (NRE)
Dr. Dingkang Zhang (NRE)
Dr. Daniel Gill (NNL)


The Continuous-Energy Coarse Mesh Transport (COMET) method is a neutron transport solution method that uses a unique hybrid stochastic-deterministic solution method to obtain high-fidelity whole-core solutions to reactor physics problems with formidable speed. This method involves pre-computing solutions to individual coarse meshes within the global problem, then using a deterministic transport sweep to construct a whole-core solution from these local solutions. In this work, a new implementation of the deterministic transport sweep solver is written which includes the ability to accelerate the calculation using up to 4 Graphics Processing Units (GPUs) on one computational node. The new implementation is written in C++ with GPU-facing logic leveraging the CUDA API. To demonstrate the new implementation, three whole-core benchmark problems were solved using the previous serial solver and various configurations of the new solver, with the relative performance compared. In this comparison, it was found that the application of one GPU to the problem resulted in between a 100x-150x speedup (depending on the specific problem) relative to the serial solver. Excellent scaling up to 4 GPUs was observed, which brought the total speedup up to 450x-500x. Although the magnitude of the speedup was found to be problem dependent, it is noted that the overall strategy of the acceleration is not problem dependent. As an example of a new type of analysis which is enabled by the improved speed of the solver, a sensitivity study was performed on the convergence thresholds used in controlling the inner and outer iteration processes. The results of the sensitivity study show that, in some examined cases, converging the problem to the limits of single-precision numbers can result in pin power errors on the order of 0.1% as compared to a completely converged double-precision result.