ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2021, Vol. 58 ›› Issue (1): 98-115.doi: 10.7544/issn1000-1239.2021.20190881

• 网络技术 • 上一篇    下一篇


朱泓睿1,2,元国军1,姚成吉3, 谭光明1, 王展1, 户忠哲1,2,3,张晓扬1,2,3,安学军1   

  1. 1(中国科学院计算技术研究所 北京 100190);2(中国科学院大学 北京 100049);3(北京旷视科技有限公司 北京 100080) (
  • 出版日期: 2021-01-01
  • 基金资助: 
    中国科学院战略性先导科技专项(B类) (XDB24050200);国家自然科学基金面上项目(61972380,61702484);中国科学院计算技术研究所创新课题(20166060)

Survey on Network of Distributed Deep Learning Training

Zhu Hongrui1,2, Yuan Guojun1, Yao Chengji3, Tan Guangming1, Wang Zhan1, Hu Zhongzhe1,2,3, Zhang Xiaoyang1,2,3, An Xuejun1   

  1. 1(Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190);2(University of Chinese Academy of Sciences, Beijing 100049);3(Megvii Inc., Beijing 100080)
  • Online: 2021-01-01
  • Supported by: 
    This work was supported by the CAS Strategic Priority Program(B) (XDB24050200), the General Program of the National Natural Science Foundation of China (61972380, 61702484), and the Innovation Fund from the Institute of Computing Technology, Chinese Academy of Sciences (20166060).

摘要: 近年来深度学习在图像、语音、自然语言处理等诸多领域得到广泛应用,但随着人们对深度学习的训练速度和数据处理能力的需求不断提升,传统的基于单机的训练过程愈发难以满足要求,分布式的深度学习训练方法成为持续提升算力的有效途径.其中训练过程中节点间网络的通信性能至关重要,直接影响训练性能.分析了分布式深度学习中的性能瓶颈,在此基础上对目前常用的网络性能优化方案进行综述,详细阐述了目前最新的超大规模分布式训练的体系结构、优化方法、训练环境和最有效的优化方法,最后对分布式训练仍然存在的困难进行了总结,对其未来研究方向进行了展望.

关键词: 分布式计算, 深度学习, 通信网络, 性能优化, 集合通信, 集群网络

Abstract: In recent years, deep learning has achieved better results than traditional algorithms in many fields such as image, speech, and natural language processing. People are increasingly demanding training speed and data processing capabilities for deep learning. However, the calculating ability of a single server has a limit and cannot achieve human demands. Distributed deep learning training has become the most effective method to expand deep learning training computing ability. At present, distributed deep learning faces a training bottleneck due to communication problems in the network during the training process which leads the communication network to be the most influential factor. There are currently many network performance optimization researches for distributed deep learning. In this paper, the main performance bottlenecks and optimization schemes are firstly demonstrated. Then the current state-of-art ultra-large-scale distributed training architecture and methods for optimization performance are specifically analyzed. Finally, a comparative summary of each performance optimization scheme and the difficulties still existing in distributed deep learning training are given, and the future research directions are pointed out as well.

Key words: distributed calculating, deep learning, communication network, performance optimization, collective communication, cluster network