• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 高性能计算 • 上一篇    下一篇

基于训练数据动态分配的深度学习并行优化机制

颜子杰,陈孟强,吴维刚   

  1. (中山大学数据科学与计算机学院,广东 广州 510006)
  • 收稿日期:2018-07-13 修回日期:2018-09-20 出版日期:2018-11-26 发布日期:2018-11-25
  • 基金资助:

    国家重点研发计划(2016YFB0200404);国家自然科学基金(U1711263)

Deep learning parallel optimization mechanism
based on dynamic distribution of training data

YAN Zijie,CHEN Mengqiang,WU Weigang   

  1. (School of Data and Computer Science,Sun Yatsen University,Guangzhou 510006,China)
  • Received:2018-07-13 Revised:2018-09-20 Online:2018-11-26 Published:2018-11-25

摘要:

基于MXNet框架,针对同步并行下参数同步耗时过长这一问题,提出了一种多机同步并行下的训练数据动态分配算法。基于计算节点的计算效率,每一次迭代后将动态调整节点需要处理的样本数据量。这样的机制使模型既能同步并行也降低了等待梯度更新的耗时。最后,利用天河二号超级计算机对此优化算法进行了对比实验,实验结果表明,所提出的优化机制达到了预期效果。

关键词: 深度学习, 数据分配, 同步并行, 并行训练, 超级计算

Abstract:

To solve the timeconsuming problem of collecting gradient updates under synchronous parallel training, we present a dynamic training data distribution algorithm under parallel synchronization of multiple machines. By calculating the computational efficiency of nodes, the amount of sample data that needs to be processed by nodes is dynamically assigned after each round of iteration. Such a mechanism allows the model to parallelize synchronously and reduce the waiting time it takes for gradient update. Finally, the mechanism is implemented via MXNet and evaluated at Tianhe2 supercomputers. Experimental results show that the proposed optimization mechanism achieves expected results.
 

Key words: deep learning, data assignment, synchronous parallel, parallel training, supercomputing