Communication-Efficient Local Stochastic Gradient Descent for Scalable Deep Learning
BigData, 2020
This paper proposes a hierarchical local-SGD training strategy that groups workers into multiple data-parallel subgroups with periodic cross-group model averaging (using a cheaper-than-allreduce averaging scheme), plus a practical metric for choosing the largest worker count that avoids significant accuracy loss, and shows experimentally that this achieves markedly better scalability than synchronous SGD while retaining comparable accuracy.
BibTeX
@inproceedings{lee2020communication,
title={Communication-efficient local stochastic gradient descent for scalable deep learning},
author={Lee, Sunwoo and Kang, Qiao and Agrawal, Ankit and Choudhary, Alok and Liao, Wei-keng},
booktitle={2020 IEEE International Conference on Big Data (Big Data)},
pages={718--727},
year={2020},
organization={IEEE}
}