DoubleSqueeze: Parallel Stochastic Gradient Descent with Double-pass Error-Compensated Compression
Hanlin Tang, Chen Yu, Xiangru Lian, Tong Zhang, Ji Liu
Abstract
A standard approach in large scale machine learning is distributed stochastic gradient training, which requires the computation of aggregated stochastic gradients over multiple nodes on a network. Communication is a major bottleneck in such applications, and in recent years, compressed stochastic gradient methods such as QSGD (quantized SGD) and sparse SGD have been proposed to reduce communication. It was also shown that error compensation can be combined with compression to achieve better convergence in a scheme that each node compresses its local stochastic gradient and broadcast the result to all other nodes over the network in a single pass. However, such a single pass broadcast approach is not realistic in many practical implementations. For example, under the popular parameter-server model for distributed learning, the worker nodes need to send the compressed local gradients to the parameter server, which performs the aggregation. The parameter server has to compress the aggregated stochastic gradient again before sending it back to the worker nodes. In this work, we provide a detailed analysis on this two-pass communication model, with error-compensated compression both on the worker nodes and on the parameter server. We show that the error-compensated stochastic gradient algorithm admits three very nice properties: 1) it is compatible with an
BibTeX
@InProceedings{pmlr-v97-tang19d,
title = {$\texttt{DoubleSqueeze}$: Parallel Stochastic Gradient Descent with Double-pass Error-Compensated Compression},
author = {Tang, Hanlin and Yu, Chen and Lian, Xiangru and Zhang, Tong and Liu, Ji},
booktitle = {Proceedings of the 36th International Conference on Machine Learning},
pages = {6155--6165},
year = {2019},
editor = {Chaudhuri, Kamalika and Salakhutdinov, Ruslan},
volume = {97},
series = {Proceedings of Machine Learning Research},
month = {09--15 Jun},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v97/tang19d/tang19d.pdf},
url = {https://proceedings.mlr.press/v97/tang19d.html},
abstract = {A standard approach in large scale machine learning is distributed stochastic gradient training, which requires the computation of aggregated stochastic gradients over multiple nodes on a network. Communication is a major bottleneck in such applications, and in recent years, compressed stochastic gradient methods such as QSGD (quantized SGD) and sparse SGD have been proposed to reduce communication. It was also shown that error compensation can be combined with compression to achieve better convergence in a scheme that each node compresses its local stochastic gradient and broadcast the result to all other nodes over the network in a single pass. However, such a single pass broadcast approach is not realistic in many practical implementations. For example, under the popular parameter-server model for distributed learning, the worker nodes need to send the compressed local gradients to the parameter server, which performs the aggregation. The parameter server has to compress the aggregated stochastic gradient again before sending it back to the worker nodes. In this work, we provide a detailed analysis on this two-pass communication model, with error-compensated compression both on the worker nodes and on the parameter server. We show that the error-compensated stochastic gradient algorithm admits three very nice properties: 1) it is compatible with an