Technical Papers Archive

Fluid-Structure Interaction Simulation of Nuclear Reactors with Ultra-Large-Scale Meshes on GPU Clusters


Authors: Xue Miao (Computer Network Information Center, Chinese Academy of Sciences); Jue Wang and Qida Lin (Computer Network Information Center, Chinese Academy of Sciences; University of Chinese Academy of Sciences); He Bai and Meng Wan (Computer Network Information Center, Chinese Academy of Sciences); Shufei Zhang (Computer Network Information Center, Chinese Academy of Sciences; University of Chinese Academy of Sciences); Kai Ren (China University of Mining and Technology Beijing); and Yangang Wang, Chunbao Zhou, Ningming Nie, and Rongqiang Cao (Computer Network Information Center, Chinese Academy of Sciences; University of Chinese Academy of Sciences)

Abstract: To overcome mapping bottlenecks and resource underutilization in large-scale nuclear reactor fluid-structure interaction simulations on heterogeneous supercomputers, we propose a GPU-accelerated framework. A matrix-free treecode (MFTC)-RBF strategy is designed to address mapping bottlenecks. It mitigates network latency and synchronization overhead caused by cross-node irregular communication via asynchronous computation-communication overlap. Furthermore, it resolves GPU instruction pipeline stalls induced by transcendental functions using a spatially-decoupled mixed-precision scheme combined with Horner polynomial approximation. To maximize intra-SM resource utilization and throughput, a dual-pipeline asynchronous scheduling (DAS) strategy is adopted to concurrently execute compute-bound and memory-bound tasks on Tensor and CUDA Cores. Evaluations show that MFTC-RBF delivers a 7.08x mapping speedup with scalability across 32,768 GPUs, while DAS hides approximately one-third of DGeMM latency. MFTC-RBF supports state-of-the-art mesh mapping with over 40-billion mesh points. End-to-end benchmarks further demonstrate a ~70% reduction in total runtime. Simulation results closely align with reference data in lead-cooled reactor benchmarks.


Presentation: file

Back to Technical Papers Archive Listing