Optimizing the spectral element CFD solver on Sunway TaihuLight for nuclear reactor simulation

Abstract

High-fidelity computational fluid dynamics (CFD) plays a crucial role in analyzing thermal–hydraulic phenomena in advanced nuclear reactors. This study presents an optimization of the spectral element method (SEM)-based CFD solver Phiflow-Solver on Sunway TaihuLight supercomputer to accelerate nuclear reactor simulations. The SEM solver relies on small, dense matrix multiplications and the Poisson operator, which are computationally challenging on heterogeneous architectures. To address these challenges, we propose two optimization strategies: (1) Porting matrix operations to the SW26010 processor’s Computing Processing Elements (CPEs) using DMA-enhanced data transfer and SIMD vectorization, achieving a 51.9% performance improvement at 64 CGs for a polynomial order of 24; (2) Enabling collaborative Management Processing Element (MPE)-CPE parallelism to compute multiple spectral elements simultaneously, achieving a 65.5% performance gain under identical conditions. By integrating these strategies, we achieve an overall 70.6% performance enhancement. Validation with a 7-pin wire-wrapped fuel assembly confirms that the heterogeneous optimizations maintain the solver’s accuracy. Furthermore, as the mesh size scales from 42 million to 1.3 billion grid points, the weak scalability remains above 90%, demonstrating the solver’s improved capability for high-resolution nuclear fuel assembly simulations.