Computational Science
Parallel Computing
Increased processing power enables high-fidelity simulation of nuclear reactor with heavy computational burden such as direct whole core calculation. The growth of processing power was driven by an increase in clock speed and an increase in the number of cores. Parallel computing is required to make the most of the performance of a computer with many cores. Many calculations are carried out simultaneously in parallel computing. One large problem can be divided into smaller ones and small problems are solved at the same time. The efficiency of parallel computation can vary greatly depending on how problems is divided, so appropriate splitting algorithm is required. Codes developed by SNURPL efficiently applied parallel computing to get the optimized performance.
Domain Decomposition
Domain decomposition is one of the frequently used splitting method. It divides one large domain into many subdomains. The small problems in subdomains are independent and appropriate for parallel computing. However the convergence of the calculation could be slow down due to the assumption for the boundary between two subdomains. Also each subdomain should has similar computational burden to get the maximum parallel performance. SNURPL has researched optimization of domain decomposition taking into account the characteristics of the target problem.
GPU Parallel Computing
Computing power growth of CPU is up to the limit nowadays. Clock speed up is limited by thermal constraints, and adding core is limited by power usage. Meanwhile GPU, which was designed for graphics rendering, devotes more to processing power rather than flow control unlike CPU. Therefore GPU is more power-effective than CPU and GPUs are increasingly becoming the main means of calculation in the scientific fields. SNURPL is playing a pioneering role in employing GPUs in the reactor physics calculations. GPU parallel computing is applied to direct whole core calculation code nTRACER and continuous energy Monte Carlo code PRAGMA. And SNURPL build its own GPU cluster server, soochiro 4, and participates as a member of high-performance computing consortium.
soochrio 4 (GPU cluster)
