On the Discrepancy between the Theoretical Analysis and Practical Implementations of Compressed Communication for Distributed Deep Learning

TL;DR

This paper investigates the discrepancy between theoretical assumptions and practical implementations of compressed communication in distributed deep learning, revealing that layer-wise compression can be theoretically superior but may not always outperform entire-model compression in practice.

Contribution

The paper proves that layer-wise compression has a better theoretical convergence bound than entire-model compression and provides an experimental comparison of six methods highlighting practical differences.

Findings

Layer-wise compression has a better theoretical convergence rate.

Practical performance varies depending on model and compression ratio.

Including both compression methods in frameworks is recommended.

Abstract

Compressed communication, in the form of sparsification or quantization of stochastic gradients, is employed to reduce communication costs in distributed data-parallel training of deep neural networks. However, there exists a discrepancy between theory and practice: while theoretical analysis of most existing compression methods assumes compression is applied to the gradients of the entire model, many practical implementations operate individually on the gradients of each layer of the model. In this paper, we prove that layer-wise compression is, in theory, better, because the convergence rate is upper bounded by that of entire-model compression for a wide range of biased and unbiased compression methods. However, despite the theoretical bound, our experimental study of six well-known methods shows that convergence, in practice, may or may not be better, depending on the actual trained model and compression ratio. Our findings suggest that it would be advantageous for deep learning frameworks to include support for both layer-wise and entire-model compression.