大规模结构网格数据的相关性统计建模轻量化方法  

Correlation Statistical Modeling Reduction Method for Large-Scale Structural Grid Data

作　　者：杨阳 武昱 汪云海[2] 曹轶[1,3] Yang Yang;Wu Yu;Wang Yunhai;Cao Yi(Institute of Applied Physics and Computational Mathematics,Beijing 100094;School of Computer Science and Technology,Shandong University,Qingdao,Shandong 266237;CAEP Software Center for High Performance Numerical Simulation,Beijing 100088)

机构地区：[1]北京应用物理与计算数学研究所,北京100094 [2]山东大学计算机科学与技术学院,山东青岛266237 [3]中物院高性能数值模拟软件中心,北京100088

出　　处：《计算机研究与发展》2023年第3期676-689,共14页Journal of Computer Research and Development

基　　金：中国博士后科学基金项目(2021M700016)。

摘　　要：高置信度的数据可视分析对于大规模数值模拟至关重要,但是当前高性能计算机的存储瓶颈导致可视分析应用获取原始高分辨率网格数据越来越困难.基于统计建模的方法能够极大降低高分辨数据存储成本,但是重建数据的不确定性高.为此,提出了一种大规模结构网格数据的相关性统计建模轻量化方法,用于对并行数值模拟生成的大规模多块体数据进行高效分析与可视化.该方法的技术核心是通过数据块间的统计相关性,指导邻接数据块的统计建模,从而有效地保留数据统计特征,且不需要对不同并行计算节点中的数据块进行合并与重新分块.通过耦合数据块的数值分布信息、空间分布信息和相关性信息,该方法可以更精确地重建原始数据,降低可视化的不确定性.实验测试采用了最大10亿网格规模的5组科学数据,定量分析结果显示,在相同数据压缩比下,该方法相比现有方法可将数据重建精度最大提升近2个数量级.Data visual analysis is essential for large-scale numerical simulations.The storage bottleneck of highperformance computers makes it challenging to analyze and visualize data with original high-resolution.The method based on statistical modeling can significantly reduce the data storage cost,with the reconstruction uncertainty being high.Therefore,we propose a large-scale data reduction method for efficient analysis and visualizing large-scale multiblock volume data generated by massively parallel scientific simulations.The technical core of this method is to guide the statistical modeling of adjacent data blocks through the statistical representation of correlation between data blocks.By doing so,our method efficiently preserves the statistical data properties without merging data blocks stored in different parallel computing nodes and repartitioning them according to the homogeneity requirements of the visualization.Compared with exsiting methods,the original data can be reconstructed more accurately by coupling numerical distribution information,spatial distribution information,and correlation information,further reducing the visual uncertainty.The experimental tests use five sets of scientific data with the largest scale of one billion grids.The quantitative analysis results show that our method improves the data reconstruction accuracy by up to two orders of magnitude at the same data compression ratio compared with the current state-of-the-art methods.

关 键 词：数据轻量化 大规模并行科学模拟 大规模多块体数据 相关性统计建模 科学可视化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]