Integrated Model, Batch and Domain Parallelism in Training Neural Networks

TL;DR

This paper introduces an integrated parallelism approach combining model, batch, and domain parallelism for efficient training of deep neural networks on large distributed systems, reducing communication costs.

Contribution

It presents a novel matrix-based integrated parallelism method that automatically combines different parallelism strategies and extends the scalability of batch parallel training.

Findings

Lower communication costs compared to pure parallelism methods

Enhanced scalability of neural network training

Analytical demonstration of efficiency improvements

Abstract

We propose a new integrated method of exploiting model, batch and domain parallelism for the training of deep neural networks (DNNs) on large distributed-memory computers using minibatch stochastic gradient descent (SGD). Our goal is to find an efficient parallelization strategy for a fixed batch size using $P$ processes. Our method is inspired by the communication-avoiding algorithms in numerical linear algebra. We see $P$ processes as logically divided into a $P_r \times P_c$ grid where the $P_r$ dimension is implicitly responsible for model/domain parallelism and the $P_c$ dimension is implicitly responsible for batch parallelism. In practice, the integrated matrix-based parallel algorithm encapsulates these types of parallelism automatically. We analyze the communication complexity and analytically demonstrate that the lowest communication costs are often achieved neither with pure model nor with pure data parallelism. We also show how the domain parallel approach can help in extending the theoretical scaling limit of the typical batch parallel method.