文章摘要
马树青,蓝强,朱勋江,邱伟,颜冰.基于水声模型多层次并行计算的匹配反演典型应用研究[J].声学技术,2023,42(5):588~593
基于水声模型多层次并行计算的匹配反演典型应用研究
Typical application study of matching inversion based on multi-level parallel computation of underwater acoustic model
投稿时间:2022-06-07  修订日期:2022-08-31
DOI:10.16300/j.cnki.1000-3630.2023.05.005
中文关键词: 海洋环境参数  匹配反演  粒子滤波  贝叶斯定理  多层次并行  天河2号  并行效率
英文关键词: ocean environmental parameters  matching inversion  particle filtering  Bayes theorem  multi-level parallelization  Tianhe Ⅱ  parallelism efficiency
基金项目:国防科技重点实验室基金(6412214200403);国防基础科研计划项目(JCKY2020550C011)。
作者单位E-mail
马树青 国防科技大学气象海洋学院, 湖南长沙 410073 mashuqing@nudt.edu.cn 
蓝强 国防科技大学气象海洋学院, 湖南长沙 410073  
朱勋江 中国人民解放军 96863部队, 河南洛阳 471000  
邱伟 国防科技大学气象海洋学院, 湖南长沙 410073  
颜冰 国防科技大学气象海洋学院, 湖南长沙 410073  
摘要点击次数: 269
全文下载次数: 244
中文摘要:
      海洋环境参数的时变特性给海洋环境参数进行直接测量带来了很多挑战,目前主流的方法是通过反演获得海洋环境参数。粒子滤波是一种重要的匹配反演方法。该方法基于贝叶斯定理,通过对海洋环境参数充分采样,能够较精确地计算出环境参数的后验概率,从而取得较好的反演性能。然而粒子滤波匹配反演方法随着粒子数的增大,计算量也急剧增大,为此,文章提出了基于水声模型的多层次并行方法,能够将粒子滤波算法高效映射到多核集群的硬件体系结构中。最后在天河2号超级计算平台进行了粒子滤波算法的并行性能测试,在单节点多核并行测试中取得了87.5%的并行效率,在多节点强扩展测试中,粒子数达到12 288个,在128个计算节点中取得了近110倍的加速性能。
英文摘要:
      Due to the time-varying nature of marine environmental parameters, there are many challenges in direct measurement of marine environmental parameters. At present, the mainstream method obtaining marine environmental parameters is inversion. Particle filtering is an important matching inversion method based on Bayes' theorem. After fully sampling marine environmental parameters, the posterior probability of environmental parameters can be calculated more accurately, so that better performance of inversion can be achieved. However, the computation complexity of particle filter matching inversion method increases rapidly with the increase of the number of particles. For this reason, a multi-level parallel computation method of the underwater acoustic model is proposed in this paper, which can efficiently map the particle filter algorithm to the hardware of the multi-core cluster machine. Finally, the parallelism performance test of the particle filter algorithm is carried out on the Tianhe Ⅱ supercomputing platform, and the parallelism efficiency of 87.5% is achieved in the single-node multi-core parallelism test. In the multi-node strong scalability test, the number of particles reaches 12 288, nearly 110 times acceleration performance is achieved in 128 computational nodes.
查看全文   查看/发表评论  下载PDF阅读器
关闭