Faster Neural Network Training with Data Echoing

TL;DR

This paper introduces data echoing, a technique that reuses intermediate outputs in the data pipeline to reduce upstream computation and accelerate neural network training, especially as hardware accelerators improve.

Contribution

We propose data echoing, a novel method that reuses data pipeline outputs to speed up training, addressing bottlenecks in data preprocessing stages.

Findings

At least one data echoing algorithm matches baseline performance with less upstream computation.

Achieved a 3.25x reduction in training time for ResNet-50 on ImageNet.

Data echoing is effective across various workloads and batch sizes.

Abstract

In the twilight of Moore's law, GPUs and other specialized hardware accelerators have dramatically sped up neural network training. However, earlier stages of the training pipeline, such as disk I/O and data preprocessing, do not run on accelerators. As accelerators continue to improve, these earlier stages will increasingly become the bottleneck. In this paper, we introduce "data echoing," which reduces the total computation used by earlier pipeline stages and speeds up training whenever computation upstream from accelerators dominates the training time. Data echoing reuses (or "echoes") intermediate outputs from earlier pipeline stages in order to reclaim idle capacity. We investigate the behavior of different data echoing algorithms on various workloads, for various amounts of echoing, and for various batch sizes. We find that in all settings, at least one data echoing algorithm can match the baseline's predictive performance using less upstream computation. We measured a factor of 3.25 decrease in wall-clock time for ResNet-50 on ImageNet when reading training data over a network.