Asynchronous Stochastic Gradient Descent with Decoupled Backpropagation and Layer-Wise Updates

TL;DR

This paper introduces PD-ASGD, an extension of asynchronous SGD that decouples forward and backward passes with layer-wise updates, significantly improving training speed and robustness in distributed deep learning.

Contribution

It proposes a novel layer-wise, decoupled ASGD method that enhances throughput and reduces delays, with theoretical convergence guarantees.

Findings

Achieves up to 5.95x faster training than synchronous methods with delays.

Runs up to 2.14x faster than existing ASGD algorithms.

Provides theoretical analysis and convergence proof for the proposed method.

Abstract

The increasing size of deep learning models has made distributed training across multiple devices essential. However, current methods such as distributed data-parallel training suffer from large communication and synchronization overheads when training across devices, leading to longer training times as a result of suboptimal hardware utilization. Asynchronous stochastic gradient descent (ASGD) methods can improve training speed, but are sensitive to delays due to both communication and differences throughput. Moreover, the backpropagation algorithm used within ASGD workers is bottlenecked by the interlocking between its forward and backward passes. Current methods also do not take advantage of the large differences in the computation required for the forward and backward passes. Therefore, we propose an extension to ASGD called Partial Decoupled ASGD (PD-ASGD) that addresses these issues. PD-ASGD uses separate threads for the forward and backward passes, decoupling the updates and allowing for a higher ratio of forward to backward threads than the usual 1:1 ratio, leading to higher throughput. PD-ASGD also performs layer-wise (partial) model updates concurrently across multiple threads. This reduces parameter staleness and consequently improves robustness to delays. Our approach yields close to state-of-the-art results while running up to $5.95\times$ faster than synchronous data parallelism in the presence of delays, and up to $2.14\times$ times faster than comparable ASGD algorithms by achieving higher model flops utilization. We mathematically describe the gradient bias introduced by our method, establish an upper bound, and prove convergence.