Optimization of neutron tracking algorithms for GPU-based continuous energy Monte Carlo calculation

• Published in: Annals of nuclear energy Vol. 162; p. 108508
• Main Authors: Choi, Namjae, Kim, Kyung Min, Joo, Han Gyu
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2021
• Subjects: Energy sort; Event-based tracking algorithm; Hashed unionized grid method; Mixed precision; Modified history-based tracking algorithm; Region partitioning; Event-based tracking algorithm; Hashed unionized grid method; Modified history-based tracking algorithm; Energy sort; Region partitioning; Mixed precision
• ISSN: 0306-4549; 1873-2100
• DOI: 10.1016/j.anucene.2021.108508

•Details of the implementations and optimizations of the GPU-based continuous-energy Monte Carlo code PRAGMA is presented.•Depleted fuel calculation is especially efficient on GPU by the vectorized cross section calculation.•Whole-core continuous-energy Monte Carlo calculation has become feasible with the GPU acceleration. The implementation and optimization strategies of the neutron tracking algorithms in the GPU-based continuous-energy Monte Carlo (MC) code PRAGMA are presented. The strategies consist of 1) a mixed precision technique, 2) an event-based tracking algorithm, 3) the history-based tracking algorithm improved for GPUs by imposing transition limits, 4) the region partitioning and energy sort schemes to vectorize macroscopic cross section calculations, 5) the unionized grid method improved by a linear hashing scheme, and 6) the GPU-specific implementation techniques such as the use of built-in vector types and an array-based race-free bank algorithm utilizing GPU atomic addition. The developed algorithms and strategies are examined for pin cell and full-core problems consisting of fresh and depleted fuels. PRAGMA turned out to be especially efficient in depleted fuel calculation; PRAGMA effectively overcame the performance drop in the depleted fuel calculation by vectorized cross section calculation. In the full-core calculations PRAGMA also rendered outstanding performance; calculations employing 11.5 billion histories and 20 consumer-grade GPUs could be finished within 15 min for a 3D APR1400 fresh core problem, and within half an hour for a mock-up depleted core problem.