GPU-based parallel Krylov linear system solver for CMFD calculation in nTRACER

The feasibility of GPU acceleration of MOC calculation in nTRACER had been explored, and the result was successful. GPU acceleration of CMFD calculation in nTRACER is required to keep up with enhanced performance of GPU-based MOC. For the solution of the large sparse linear system involved in the CMFD formulation, the preconditioned Krylov iterative solver, especially preconditioned BiConjugate Gradient Stabilized solver (pBiCGSTAB), has been used as the linear solver of nTRACER. Therefore GPU-based pBiCGSTAB linear system solver is constructed in this work. GPU-based pBiCGSTAB solver is constructed to accelerate CMFD calculation in nTRACER. The evaluation of spMv algorithms in GPU shows sellp algorithm which allows fully coalescent memory access outperforms other algorithms. Fully coalescent memory access maximizes data transfer per cycle on GPU. In addition, explicit solution of whole linear system shows better performance than Gauss-Seidel group sweep on GPU due to the overhead cost caused by communication between CPU and GPU. Both sellp and explicit solution increases the total arithmetic cost when compared to conventional algorithms. These results show that the performance of Krylov iterative solver in parallel heterogeneous system is determined by communication cost rather than arithmetic cost.