Revisiting Distributed Synchronous SGD

TL;DR

This paper reevaluates distributed training methods for deep learning, showing that synchronous optimization with backup workers can outperform traditional asynchronous and synchronous approaches in speed and accuracy.

Contribution

It introduces a hybrid approach using backup workers in synchronous SGD, reducing noise and improving convergence and test accuracy.

Findings

Synchronous SGD with backup workers converges faster.

The approach achieves better test accuracies.

It mitigates straggler effects effectively.

Abstract

Distributed training of deep learning models on large-scale training data is typically conducted with asynchronous stochastic optimization to maximize the rate of updates, at the cost of additional noise introduced from asynchrony. In contrast, the synchronous approach is often thought to be impractical due to idle time wasted on waiting for straggling workers. We revisit these conventional beliefs in this paper, and examine the weaknesses of both approaches. We demonstrate that a third approach, synchronous optimization with backup workers, can avoid asynchronous noise while mitigating for the worst stragglers. Our approach is empirically validated and shown to converge faster and to better test accuracies.