Computational methods for efficient nuclear data management in Monte Carlo neutron transport simulations

Abstract

This thesis presents the development and analysis of computational methods for efficiently accessing and utilizing nuclear data in Monte Carlo neutron transport code simulations. Using the OpenMC code, profiling studies are conducted in order to determine the types of nuclear data that are used in realistic reactor physics simulations, as well as the frequencies with which those data are accessed. The results of the profiling studies are then used to motivate the conceptualization of a nuclear data server algorithm aimed at reducing on-node memory requirements through the use of dedicated server nodes for the storage of infrequently accessed data. A communication model for this algorithm is derived and used to make performance predictions given data access frequencies and assumed system hardware parameters. Additionally, a new, accelerated approach for rejection sampling the free gas resonance elastic scattering kernel that reduces the frequency of zero-temperature elastic scattering cross section data accesses is derived and implemented. Using this new approach, the runtime overhead incurred by an exact treatment of the free gas resonance elastic scattering kernel is reduced by more than 30% relative to a standard sampling procedure used by Monte Carlo codes. Finally, various optimizations of the commonly-used binary energy grid search algorithm are developed and demonstrated. Investigated techniques include placing kinematic constraints on the range of the searchable energy grid, index lookups on unionized material energy grids, and employing energy grid hash tables. The accelerations presented routinely result in overall code speedup by factors of 1.2-1.3 for simulations of practical systems.