面向科学计算可视化的两级并行数据读取加速方法

发布时间：2018-05-20 07:41

本文选题：两级并行数据读取 + 科学计算可视化　；参考：《计算机研究与发展》2017年04期

【摘要】：为了匹配超级计算机的整体计算能力,超级计算机存储子系统通常具有良好的I/O性能可扩展性,表现为:应用获得存储子系统最佳性能时的I/O访问并发度,与超级计算机系统总计算核数(可达数万至数百万)通常处于同一数量级.然而,科学计算可视化应用通常使用的进程数(等于I/O访问并发度)相对较小(经验上常设为计算进程数的1%,典型值为数个至数百个),因此无法充分发挥超级计算机存储子系统的最佳I/O性能.提出了一种面向科学计算可视化的两级并行数据读取加速方法,在可视化进程内部引入多线程并行数据读取,通过进程间和进程内两级并行,增加超级计算机存储子系统的I/O访问并发度,提升可视化应用数据读取速率.测试结果表明:在不同的可视化进程规模下,两级并行比单级并行峰值数据读取速率提高33.5%~269.5%,均值数据读取速率提高26.6%~232.2%;随着科学计算应用种类以及应用规模的变化,两级并行数据读取可使可视化应用整体峰值运行速度加速19.5%~225.7%,均值运行速度加速15.8%~197.6%.
[Abstract]:In order to match the overall computing power of the supercomputer, the storage subsystem of the supercomputer usually has good I / O performance scalability. It is usually in the same order of magnitude as the total count of supercomputer systems (up to tens of thousands of millions). However, The number of processes (equal to I / O access concurrency) usually used in visualization applications for scientific computing is relatively small (empirically permanent is the number of processes calculated, typical values range from several to hundreds), so it is not possible to make full use of supercomputer storage. Optimal I / O performance of the storage subsystem. In this paper, a two-level parallel data reading acceleration method for scientific computation visualization is proposed. Multithread parallel data reading is introduced into the visualization process. Increases the I / O access concurrency of the supercomputer storage subsystem and improves the data reading rate of visual applications. The test results show that under different visualization process scales, the reading rate of two-level parallel data is 33.5269.5% higher than that of single-level parallel peak data, and the average data reading rate is 26.6% and 232.2% higher than that of single-level parallel peak data. The two-level parallel data reading can accelerate the whole peak speed of visual application by 19.5and 225.7. the average speed can be accelerated by 15.8and 197.6.
【作者单位】：中国舰船研究院;北京应用物理与计算数学研究所;中物院高性能数值模拟软件中心;
【分类号】：TP391

【相似文献】