摘要： 为满足大规模机器学习系统高定制化、低耦合与低资源消耗的需求，设计并实现一个轻量级分布式机器学习系统。采用模块化分层设计并移植多种主流的机器学习与深度学习算法，同时提出参数服务器与动态Ring-AllReduce 2种可扩展梯度同步方案，对算法模型进行并行训练加速。实验结果表明，该系统对于稀疏与稠密模型均有较好的扩展性与稳定性，参数服务器训练可达到与单机相近的准确率与收敛效果，Ring-AllReduce也能在8节点模型上实现相对单节点模型6倍的训练加速。

关键词: 机器学习系统, 分布式系统, 并行计算, 集合通信, 模块化

Abstract: To improve customization and reduce coupling and resource consumption in large-scale machine learning systems,this paper designs and realizes a lightweight distributed machine learning system with high performance and scalability.The system adopts a modular and layered design,and migrates a variety of mainstream machine learning and deep learning algorithms.Two kinds of efficient extensible gradient synchronization schemes,parameter server and Ring All-Reduce,are proposed to perform parallel training acceleration experiments on the mainstream algorithm models.Experimental results show that the system can achieve excellent scalability and stability for both sparse and dense models.Parameter Server training can achieve similar accuracy and convergence performance to those of standalone training.The proposed Ring All-Reduce can achieve a 6x training acceleration on 8-node model compared with single node model.

Key words: machine learning system, distributed system, parallel computing, collective communication, modularity

中图分类号: 

• TP18