| 郭卓强* **,毛润泽***,谭光明*,陈帜****,贾伟乐*.基于人工智能的高精度超临界燃烧模拟在众核超算上的优化[J].高技术通讯(中文),2026,36(5):455~466 |
| 基于人工智能的高精度超临界燃烧模拟在众核超算上的优化 |
| Optimization of artificial intelligence-based high-accuracy supercritical combustion simulation on many-core supercomputers |
| |
| DOI：10. 3772 / j. issn. 1002 - 0470. 2026. 05. 002 |
| 中文关键词: 高性能计算； 深度神经网络； 燃烧模拟； 计算流体力学 |
| 英文关键词: high-performance computing, deep neural network, flame simulation, computational fluid dynamics |
| 基金项目: |
| 作者 | 单位 | | 郭卓强* ** | （*处理器芯片全国重点实验室，中国科学院计算技术研究所北京 100190）
（**中国科学院大学北京 100049）
（***北京大学工学院北京 100871）
（****京科学智能研究院北京 100084） | | 毛润泽*** | | | 谭光明* | | | 陈帜**** | | | 贾伟乐* | |
|
| 摘要点击次数: 23 |
| 全文下载次数: 13 |
| 中文摘要: |
| 科学智能正在重塑计算科学的研究范式，其中高精度超临界燃烧模拟面临着计算复杂度与物理精度协同优化的双重挑战。基于人工智能的燃烧模拟框架DeepFlame虽然通过算法创新有效平衡了物理精度与计算效率，但在现代长向量众核超算平台上仍在神经网络推理、偏微分方程组求解及大规模文件读写3个维度存在性能瓶颈。本研究提出系统性优化方案：首先针对神经网络推理，通过混合精度计算、算子融合及查表拟合等技术实现计算效率提升；其次针对偏微分方程组求解，创新性地提出网格重排、挑战稀疏矩阵格式、网格染色等技术，构建面向众核架构的偏微分方程并行求解器，有效改善稀疏线性系统的访存局部性差、间接访存和写冲突等问题；最后针对十万核级并行场景，设计基于多程序融合、文件索引和两级并行读写的优化机制，突破大规模模拟时初始化数据加载的性能瓶颈。实验结果表明，程序优化后在众核架构富岳超算平台上实现4.30倍加速比，在大规模模拟中浮点运算效率达到了理论峰值的31.80%。另外，本研究还实现了193亿网格规模的高保真超临界燃烧模拟，为下一代燃烧室设计提供了突破性的数值研究工具。 |
| 英文摘要: |
| Artificial intelligence (AI) for Science is reshaping the paradigms of scientific computing, where high-accuracy supercritical combustion simulation faces dual challenges of balancing computational complexity and physical fidelity. Although the AI-based combustion simulation framework DeepFlame achieves a trade-off between physical accuracy and computational efficiency through algorithmic innovations, it still encounters performance bottlenecks on modern long-vector many-core supercomputing platforms in three critical dimensions: neural network inference, partial differential equation (PDE) solving, and large-scale input/output (I/O). This study proposes a systematic optimization approach: (1) For neural network inference, computational efficiency is enhanced via mixed-precision computing, kernel fusion, and lookup table approximation; (2) For PDE solving, an innovative many-core-optimized solver is developed using mesh renumbering, custom sparse matrix formats, mesh coloring, addressing issues such as poor memory locality, indirect memory access, and write conflicts in sparse linear systems; (3) For extreme-scale parallelism (100000+ cores), a high-performance initialization mechanism is designed with multi-program fusion, file indexing, and two-level parallel I/O, overcoming data loading bottlenecks. Experimental validation on the ARM-based Fugaku supercomputer demonstrates a 4.3×speedup and a floating-point efficiency reaching 31.8% of the theoretical peak in large-scale simulations. Additionally, this work achieves high-fidelity supercritical combustion simulation at a 19.3 billion-cell scale, providing a groundbreaking numerical tool for next-generation combustor design. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
