基于机器学习的MPI仿真性能测试与调优平台

Machine Learning‑Based MPI Simulation Performance Testing and Tuning Platform

摘要: 工程数值模拟软件的MPI并行性能测试过程中，存在MPI参数调优难、人工测试低效、性能基线更新不及时等痛点，如何提升并行性能测试与调优效率成为并行计算领域研究热点之一。本文基于机器学习技术和XGBoost算法构建MPI参数与性能映射模型，运用惯性权重自适应调整、引入高斯变异策略改进IPSO自动调优算法，实现MPI并行参数自动调优，研发自动化回归测试技术，提出统计分析与异常剔除相结合的性能基线生成方法，最终设计并实现面向MPI并行软件研发的参数调优与自动化性能测试SimPerf平台。实验验证结果表明，该平台可显著降低仿真研发人员的MPI并行开发门槛、提升研发与测试效率，与传统人工方式对比，MPI并行仿真应用的性能测试与调优效率提升70%以上，测试覆盖率提升16%以上。

Abstract: During the MPI parallel performance testing of engineering numerical simulation software, several critical challenges exist, including difficult tuning of MPI parameters, low efficiency of manual testing, and untimely updates of performance baselines. Enhancing the efficiency of parallel performance testing and parameter tuning has therefore become a research hotspot in the field of parallel computing. In this paper, a mapping model between MPI parameters and performance is constructed based on machine learning techniques and the XGBoost algorithm. An improved IPSO automatic tuning algorithm is proposed by adaptively adjusting the inertia weight and introducing a Gaussian mutation strategy to realize automatic tuning of MPI parallel parameters. Automated regression testing technology is developed, and a performance baseline generation method combining statistical analysis and outlier elimination is put forward. Ultimately, the SimPerf platform for parameter tuning and automated performance testing is designed and implemented for the research and development of MPI parallel software. Experimental validation results demonstrate that the proposed platform can significantly lower the MPI parallel development threshold for simulation researchers and improve research‑development and testing efficiency. Compared with traditional manual methods, the efficiency of performance testing and tuning for MPI parallel simulation applications is improved by more than 70%, and test coverage is increased by over 16%.