Towards Scalable Distributed Training of Deep Learning on Public Cloud Clusters

TL;DR

This paper introduces a new communication-efficient method and system optimizations for distributed deep learning training on public cloud clusters, achieving significant speedups and record-breaking results.

Contribution

It proposes a top-k sparsification communication library and system-level optimizations to improve scalability and efficiency of distributed training on public cloud GPU clusters.

Findings

Achieves 25%-40% faster training than existing systems.

Breaks DAWNBench record for ResNet-50 training on ImageNet.

Demonstrates effectiveness on Tencent Cloud GPU clusters.

Abstract

Distributed training techniques have been widely deployed in large-scale deep neural networks (DNNs) training on dense-GPU clusters. However, on public cloud clusters, due to the moderate inter-connection bandwidth between instances, traditional state-of-the-art distributed training systems cannot scale well in training large-scale models. In this paper, we propose a new computing and communication efficient top-k sparsification communication library for distributed training. To further improve the system scalability, we optimize I/O by proposing a simple yet efficient multi-level data caching mechanism and optimize the update operation by introducing a novel parallel tensor operator. Experimental results on a 16-node Tencent Cloud cluster (each node with 8 Nvidia Tesla V100 GPUs) show that our system achieves 25%-40% faster than existing state-of-the-art systems on CNNs and Transformer. We finally break the record on DAWNBench on training ResNet-50 to 93% top-5 accuracy on ImageNet.