| 丁武,杨芳,王卫光,王汉岗,何用,蔺崇哲.二维水动力模型多GPU分布式数据并行计算方法研究[J].水利学报,2025,56(12):1647-1658 |
| 二维水动力模型多GPU分布式数据并行计算方法研究 |
| Research on multi-GPU distributed data parallel computing methods for 2D hydrodynamic models |
| 投稿时间：2025-05-07 修订日期：2025-12-23 |
| DOI：10.13243/j.cnki.slxb.20250262 |
| 中文关键词: 二维水动力模型 有限体积法 计算域分割 NCCL 异构计算 |
| 英文关键词: 2D hydrodynamic model finite volume method computational domain partitioning NCCL communica? tion heterogeneous parallel computing |
| 基金项目:国家重点研发计划项目（2024YFC3212000） |
|
| 摘要点击次数: 11 |
| 全文下载次数: 16 |
| 中文摘要: |
| 针对二维水动力模型在复杂流域洪水模拟中的算力瓶颈，本研究构建了基于物理拓扑分域与基于分布式数据并行及NCCL异步通信融合的多GPU异构并行架构，实现非结构三角网格水动力模型的超算级加速。通过物理拓扑保持型计算域分割算法，在保障非结构三角网格邻接关系完整性的同时，实现多GPU间的动态负载均衡；结合基于Godunov有限体积法的分布式数据并行求解框架与NCCL异步通信策略，构建了“计算-通信”协同的异构加速架构。实验验证表明：在二维理想溃坝算例中，模型能精准捕捉溃坝激波传播特征，与理论解保持高度一致。在崂山水库溃坝洪水模拟案例中，利用8块GPU并行计算实现了15.70 s完成2 h的溃坝洪水演进模拟，取得了308.43倍的加速比，其中各子域间的数据通信效率较传统MPI提升了8.67%。该架构支持单节点至跨节点GPU集群的弹性扩展，可为数字孪生流域提供秒级响应的超算级水动力引擎，为推动防洪预报调度从静态预案向动态预演范式提供核心技术支撑。 |
| 英文摘要: |
| To address the computational bottlenecks of 2D hydrodynamic models for flood simulation of complex watersheds，this study develops a multi-GPU heterogeneous parallel architecture based on physical topology-based domain partitioning and integration of distributed data parallelism with NCCL asynchronous communication，enabling supercomputing-level acceleration for unstructured triangular mesh-based hydrodynamic models. Through a physical topology-preserving domain partitioning algorithm，dynamic load balancing across multiple GPUs was achieved， while ensuring the integrity of adjacency relationships within unstructured triangular meshes. By combining a distributed data parallel solution framework based on the Godunov finite volume method with an NCCL asynchronous com munication strategy，a heterogeneous acceleration architecture with "computation-communication" collaboration was established. Experimental validation shows that，in a 2D idealized dam-break test case，the model accurately captured the propagation characteristics of the dam-break shock waves，maintaining high consistency with theoretical solutions. In the case study of the Laoshan Reservoir dam-break flood simulation，the 2-hour flood evolution simulation was completed in just 15.70 s using 8 GPUs for parallel computing，achieving a speedup of 308.43 times. Furthermore，the data communication efficiency between subdomains improved by 8.67% compared to traditional MPIbased methods. The proposed architecture supports elastic scaling from single-node to cross-node GPU clusters，providing a supercomputing-level hydrodynamic engine with second-level response capabilities for digital twin watersheds. It offers core technological support for advancing flood forecasting and dispatching from static planning toward dynamic rehearsal paradigms. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |