Compressed Communication for Distributed Training: Adaptive Methods and System

TL;DR

This paper introduces a new adaptive gradient method with gradient compression and a scalable system, BytePS-Compress, significantly reducing communication overhead and training time in distributed machine learning without accuracy loss.

Contribution

The paper proposes a novel adaptive gradient method with gradient compression and develops BytePS-Compress, a system enabling efficient two-way gradient compression in distributed training.

Findings

Improved training times for ResNet50, VGG16, and BERT-base by up to 58.1%.

Achieved a 333x compression rate for BERT training.

Maintained model accuracy despite high compression rates.

Abstract

Communication overhead severely hinders the scalability of distributed machine learning systems. Recently, there has been a growing interest in using gradient compression to reduce the communication overhead of the distributed training. However, there is little understanding of applying gradient compression to adaptive gradient methods. Moreover, its performance benefits are often limited by the non-negligible compression overhead. In this paper, we first introduce a novel adaptive gradient method with gradient compression. We show that the proposed method has a convergence rate of $\mathcal{O}(1/\sqrt{T})$ for non-convex problems. In addition, we develop a scalable system called BytePS-Compress for two-way compression, where the gradients are compressed in both directions between workers and parameter servers. BytePS-Compress pipelines the compression and decompression on CPUs and achieves a high degree of parallelism. Empirical evaluations show that we improve the training time of ResNet50, VGG16, and BERT-base by 5.0%, 58.1%, 23.3%, respectively, without any accuracy loss with 25 Gb/s networking. Furthermore, for training the BERT models, we achieve a compression rate of 333x compared to the mixed-precision training.