ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2015, Vol. 52 ›› Issue (4): 861-868.doi: 10.7544/issn1000-1239.2015.20131920

• 系统结构 • 上一篇    下一篇

基于HDF5实现多区结构网格CFD程序的并行I/O

杨丽鹏,车永刚   

  1. (国防科学技术大学并行与分布处理国防科技重点实验室 长沙 410073) (yanglipeng009@163.com)
  • 出版日期: 2015-04-01
  • 基金资助: 
    基金项目:航天支撑基金项目(2012-HT-GFKD-19);空气动力学国家重点实验室基金项目(SKLA20130105)

HDF5 Based Parallel I/O Techniques for Multi-Zone Structured Grids CFD Applications

Yang Lipeng, Che Yonggang   

  1. (Science and Technology on Parallel and Distributed Processing Laboratory, National University of Defense Technology, Changsha 410073)
  • Online: 2015-04-01

摘要: 计算流体动力学(computational fluid dynamics, CFD)是高性能计算重要应用领域之一,其计算涉及大量数据访问.在大规模并行计算情况下,串行I/O的性能与计算能力不匹配,I/O成为性能瓶颈.并行I/O是解决这一问题的主要途径之一.针对一个真实多区结构网格CFD并行程序HOSTA(high-order simulator for aerodynamics),基于HDF5(hierarchical data format v5)数据存储格式及其并行I/O编程接口,实现了其主要数据的并行I/O.在一套有6个I/O服务器结点的高性能计算机系统上,采用实际CFD算例进行了性能测试.对一个三角翼算例,并行I/O相对于串行I/O的性能加速比达到21.27,最高获得5.81GBps的I/O吞吐率,并使程序整体性能提高10%以上;对一个网格规模更大的简单翼型算例,并行I/O最高获得了6.72GBps的I/O吞吐率.

关键词: 计算流体动力学, 并行I/O, 层次式数据存储格式, 多区结构网格, I/O吞吐率

Abstract: Computational fluid dynamics (CFD) is one of the most important high performance computing (HPC) areas. CFD applications commonly access large volumes of data. In case of large-scale CFD parallel computing, serial I/O performance does not match the computing performance, hence it becomes the performance bottleneck. Parallel I/O is an effective way to solve this problem. HDF5 (hierarchical data format v5) has provided excellent mechanisms to manage scientific data as well as effective ways to implement parallel I/O. High-order simulator for aerodynamics (HOSTA) is a multi-zone structured grids CFD application that can solve real world flow problems. This paper implements the parallel I/O method in HOSTA, based on the HDF5 file format and corresponding parallel I/O application programming interface. Detailed performance evaluation is performed with real CFD simulation cases on a HPC system equipped with 6 I/O service nodes. The results show that our method is both scalable and efficient. For a delta wing test case, parallel I/O achieves the performance speedup of 21.27 compared with the serial I/O. It achieves maximal I/O throughput of 5.81GBps and improves the application’s performance by over 10%, as compared with the original code. For a simple airfoil test case with larger grid size, our parallel I/O achieves maximal I/O throughput of 6.72GBps.

Key words: computational fluid dynamics (CFD), parallel I/O, hierarchical data format v5 (HDF5), multi-zone structured grids, I/O throughput

中图分类号: