Why Dataset Properties Bound the Scalability of Parallel Machine Learning Training Algorithms

doi:10.1109/TPDS.2020.3048836

	Why Dataset Properties Bound the Scalability of Parallel Machine Learning Training Algorithms
	Cheng, Daning 1; Li, Shigang 2; Zhang, Hanping 3; Xia, Fen 3; Zhang, Yunquan 4
	2021-07-01
发表期刊	IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS
ISSN	1045-9219
卷号	32 期号:7 页码:1702-1712
摘要	As the training dataset size and the model size of machine learning increase rapidly, more computing resources are consumed to speedup the training process. However, the scalability and performance reproducibility of parallel machine learning training, which mainly uses stochastic optimization algorithms, are limited. In this paper, we demonstrate that the sample difference in the dataset plays a prominent role in the scalability of parallel machine learning algorithms. We propose to use statistical properties of dataset to measure sample differences. These properties include the variance of sample features, sample sparsity, sample diversity, and similarity in sampling sequences. We choose four types of parallel training algorithms as our research objects: (1) the asynchronous parallel SGD algorithm (Hogwild! algorithm), (2) the parallel model average SGD algorithm (minibatch SGD algorithm), (3) the decentralization optimization algorithm, and (4) the dual coordinate optimization (DADM algorithm). Our results show that the statistical properties of training datasets determine the scalability upper bound of these parallel training algorithms.
关键词	Training Scalability Machine learning Machine learning algorithms Stochastic processes Task analysis Upper bound Parallel training algorithms training dataset scalability stochastic optimization methods
DOI	10.1109/TPDS.2020.3048836
收录类别	SCI
语种	英语
资助项目	National Natural Science Foundation of China[61972376] ; National Natural Science Foundation of China[61502450] ; National Natural Science Foundation of China[61432018] ; National Natural Science Foundation of China[61521092] ; National Key Research and Development Program of China[2016YFB0200800] ; National Key Research and Development Program of China[2016YFB0200803] ; National Key Research and Development Program of China[2017YFB0202302] ; National Key Research and Development Program of China[2017YFB0202105] ; State Key Laboratory of Computer Architecture Foundation[CARCH3504] ; Natural Science Foundation of Beijing[L182053]
WOS研究方向	Computer Science ; Engineering
WOS类目	Computer Science, Theory & Methods ; Engineering, Electrical & Electronic
WOS记录号	WOS:000621405200017
出版者	IEEE COMPUTER SOC
引用统计	被引频次：9[WOS] [WOS记录] [WOS相关记录]
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/16911
专题	中国科学院计算技术研究所期刊论文_英文
通讯作者	Zhang, Yunquan
作者单位	1.Chinese Acad Sci, Inst Comp Technol, SKL, Beijing, Peoples R China 2.Swiss Fed Inst Technol, Dept Comp Sci, Zh, Switzerland 3.Beijing Wisdom Uranium Technol Co Ltd, Algorithm Dept, Beijing, Peoples R China 4.Chinese Acad Sci, Inst Comp Technol, SKL Comp Architecture, Beijing, Peoples R China
推荐引用方式 GB/T 7714	Cheng, Daning,Li, Shigang,Zhang, Hanping,et al. Why Dataset Properties Bound the Scalability of Parallel Machine Learning Training Algorithms[J]. IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS,2021,32(7):1702-1712.
APA	Cheng, Daning,Li, Shigang,Zhang, Hanping,Xia, Fen,&Zhang, Yunquan.(2021).Why Dataset Properties Bound the Scalability of Parallel Machine Learning Training Algorithms.IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS,32(7),1702-1712.
MLA	Cheng, Daning,et al."Why Dataset Properties Bound the Scalability of Parallel Machine Learning Training Algorithms".IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS 32.7(2021):1702-1712.