数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (5): 102-112.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.05.008

doi: 10.11871/jfdc.issn.2096-742X.2025.05.008

• 专刊：国产算力新力量，助力科学计算应用新发展 • 上一篇    下一篇

基于国产异构加速平台的E+A星系搜索：并行化策略与实现

郑爱宇(),孟祥宇,张博宇,周立婵,杨海峰*()   

  1. 太原科技大学，山西 太原 030024
  • 收稿日期:2025-02-25 出版日期:2025-10-20 发布日期:2025-10-23
  • 通讯作者: 杨海峰
  • 作者简介:郑爱宇，太原科技大学，博士研究生，主要研究方向为天文大数据背景下的机器学习、人工智能技术研究。
    本文承担工作为：模型分析与设计，论文撰写，模型算法设计。
    ZHENG Aiyu is a Ph.D. student at Taiyuan University of Science and Technology. His main research interests include machine learning and artificial intelligence technologies in the context of astronomical big data.
    In this paper, he is mainly responsible for model analysis and design, paper writing and model algorithm design.
    E-mail: zheng_aiyu@tyust.edu.cn|杨海峰，太原科技大学，博士生导师，主要研究方向为天文大数据背景下的数据挖掘技术。
    本文承担工作为：指导优化模型和模型设计。
    YANG Haifeng is a Ph.D. supervisor at Taiyuan University of Science and Tchnology. His main research interests include data mining techniques in the context of astronomical big data.
    In this paper, he is mainly responsible for providing the guidance for optimizing and designing models.
    E-mail: hfyang@tyust.edu.cn
  • 基金资助:
    山西省科技合作与交流项目(202204041101037);山西省科技合作与交流项目(2022040411101033);光合基金(ghfund202302032024);光合基金(ghfund202407027490)

E+A Galaxy Search Based on a Domestic Heterogeneous Acceleration Platform: Parallelization Strategy and Implementation

ZHENG Aiyu(),MENG Xiangyu,ZHANG Boyu,ZHOU Lichan,YANG Haifeng*()   

  1. Taiyuan University of Science and Technology, Taiyuan, Shanxi 030024, China
  • Received:2025-02-25 Online:2025-10-20 Published:2025-10-23
  • Contact: YANG Haifeng

RichHTML

0

PDF

25

摘要：

【目的】 E+A星系是生命周期短暂且稀有的后星爆星系，其实测样本对研究宇宙的形成与演化具有重要价值。各种巡天望远镜的持续观测积累了超海量的天文数据，而提升稀有目标的挖掘和识别效率是当前研究的挑战之一。【方法】 本文提出一种基于国产异构加速平台的并行E+A星系搜寻方法，称为Parallel E+A Searcher，PEAS。本文首先逐步骤分析了串行的E+A星系搜寻算法的逻辑依赖性和数据依赖性，确定了三个可并行化的任务算子。然后基于国产异构加速平台提供的软件栈和组件设计了PEAS的层次化的分布式并行架构。最后，我们基于平台提供的OpenMP和MPI框架分别设计了PEAS的多核任务并行方案以及多节点分布式数据并行方案。【结果】 PEAS在LAMOST DR2数据集上通过了正确性验证。在LAMOST DR10中的26万条星系数据集上开展了围绕加速比的性能测试，结果表明：与单核CPU相比，PEAS在CPU32的加速比为22.30，在单加速卡的加速比高达107.06；在性能扩展性方面，4加速卡对应1加速卡有加速1.89，4节点对应1节点有加速1.83；在数据扩展性方面，扩展加速比为6.93，逼近于数据比例8.6。

关键词: 国产异构加速平台, 并行计算, E+A星系, 稀有天文目标搜寻

Abstract:

[Objective] E+A galaxies are rare, short-lived post-starburst galaxies whose observational samples hold critical value for understanding galactic evolution and cosmic history. While modern sky surveys have amassed vast astronomical datasets, efficiently detecting these transient objects remains a key challenge in contemporary astrophysical research. [Methods] This study proposes PEAS (Parallel E+A Searcher), a novel pipeline for accelerated E+A galaxy detection implemented on a domestic heterogeneous computing platform. Our methodology involves three phases. First, we analyze dependencies in the serial search algorithm to decompose it into three parallelizable task operators. Second, leveraging the software stack of the target platform, we design a hierarchical distributed architecture for PEAS. Finally, we implement two parallelization schemes: a multi-core task-parallel approach using OpenMP and a multi-node data-parallel strategy using MPI. [Results] Validation on the LAMOST DR2 dataset confirms PEAS’s accuracy. Performance benchmarks conducted on 260,000 galaxies from LAMOST DR10 demonstrate significant speedups. The results indicate that, compared to a single-core CPU, PEAS achieves a speedup of 22.30 on a 32-core system and of up to 107.06 on a single accelerator card. In terms of performance scalability, 4 acceleration cards achieve a speedup 1.89 compared to 1 acceleration card, while 4 nodes achieve 1.83 speedup compared to 1 node. In terms of data scalability, the speedup is 6.93, approaching the data ratio of 8.6.

Key words: domestic heterogeneous acceleration platform, Parallel computing, E+A galaxy, rare astronomical target search