首页|期刊导航|水科学进展|适配国产多节点GPU的二维非结构高性能水动力模型

适配国产多节点GPU的二维非结构高性能水动力模型OA

A high-performance 2D unstructured hydrodynamic model for domestic multi-node GPU clusters

中文摘要英文摘要

为满足国产信创环境下洪水数值模拟的高性能运算要求,研发了适配国产多节点 GPU集群的二维非结构高性能水动力模型.该模型基于二维有限体积法的 Godunov型数值格式,依托可移植异构计算接口(HIP)框架,采用C++编程语言实现适配国产 GPU的高性能加速计算,为进一步提升计算效率,采用 METIS区域分解法和消息传递接口技术(MPI)实现跨节点多 GPU的高效协同计算.通过经典矩形溃坝算例及 Toce河算例验证了新研发模型具有良好的数值精度、可靠的稳定性和正确的分区通信技术,将模型在江苏省横山水库进行应用测试,结果表明:在"米"级分辨率、1 500万网格规模下,使用 16个节点 64块 GPU可将计算耗时从 3.76 h缩短至 15 min,实现了大规模、高分辨率洪水模拟由单卡"小时级"到多节点"分钟级"的效率跨越.新研发模型可完全适配国产软硬件环境,具有广泛的工程应用前景.

To meet the computational demands of numerical flood simulation under China's information technology innovation initiative,this study proposes a high-performance two-dimensional unstructured hydrodynamic model designed for domestic multi-node GPU clusters.The model is based on a Godunov-type finite volume scheme and is implemented in C++using the Heterogeneous-Compute Interface for Portability(HIP)framework,enabling efficient acceleration on domestic GPUs.METIS-based domain decomposition and the message passing interface(MPI)are adopted to support collaborative computation across multiple GPUs and computing nodes,thereby improving parallel computing efficiency.Numerical tests using the classical rectangular dam-break benchmark and the Toce River case demonstrate that the model achieves satisfactory accuracy,stable computation,and reliable inter-domain communication.Field application to Hengshan Reservoir in Jiangsu Province further shows that,for a meter-scale simulation containing 15 million mesh cells,the use of 64 GPUs across 16 nodes reduced the total runtime from 3.76 h to 15 min.This result indicates a substantial improvement in computational efficiency for large-scale,high-resolution flood simulation,reducing runtime from the hourly scale on a single GPU to the minute scale on a multi-node GPU cluster.With compatibility across domestic hardware and software environments,the proposed model has strong potential for engineering applications in high-resolution flood modeling.

张大伟;冯新政;马利平;王帆;王玮琦;孙东亚

中国水利水电科学研究院流域水循环与水安全全国重点实验室,北京 100038中国水利水电科学研究院流域水循环与水安全全国重点实验室,北京 100038西安理工大学水利水电学院,陕西 西安 710048中国水利水电科学研究院流域水循环与水安全全国重点实验室,北京 100038中国水利水电科学研究院流域水循环与水安全全国重点实验室,北京 100038中国水利水电科学研究院流域水循环与水安全全国重点实验室,北京 100038

建筑与水利

二维水动力模型国产GPU多节点非结构网格Godunov格式

2D hydrodynamic modeldomestic GPUmulti-node computingunstructured meshGodunov scheme

《水科学进展》 2026 (5)

917-925,9

国家重点研发计划项目(2024YFC3214802)中国水科院科研成果转化基金专项(JZ121003A0052025) The study is financially supported by the National Key R&D Program of China(No.2024YFC3214802)and Special Fund for Scientific Achievements Transformation of IWHR,China(No.JZ121003A0052025).

10.14042/j.cnki.32.1309.2026.05.011

评论