Efficient and Robust Parallel DNN Training through Model Parallelism on Multi-GPU Platform

TL;DR

This paper introduces a pipelined model parallel training method with a novel weight prediction technique, SpecTrain, achieving significant speedup and stable accuracy in multi-GPU DNN training.

Contribution

It proposes a new weight prediction method, SpecTrain, that enables efficient pipelined model parallelism with high GPU utilization and stable training accuracy.

Findings

Achieves up to 8.91x speedup over data parallelism.

Maintains comparable model accuracy with improved training efficiency.

Reduces inter-GPU communication costs significantly.

Abstract

The training process of Deep Neural Network (DNN) is compute-intensive, often taking days to weeks to train a DNN model. Therefore, parallel execution of DNN training on GPUs is a widely adopted approach to speed up the process nowadays. Due to the implementation simplicity, data parallelism is currently the most commonly used parallelization method. Nonetheless, data parallelism suffers from excessive inter-GPU communication overhead due to frequent weight synchronization among GPUs. Another approach is pipelined model parallelism, which partitions a DNN model among GPUs, and processes multiple mini-batches concurrently. This approach can significantly reduce inter-GPU communication cost compared to data parallelism. However, pipelined model parallelism faces the weight staleness issue; that is, gradients are computed with stale weights, leading to training instability and accuracy loss. In this paper, we present a pipelined model parallel execution method that enables high GPU utilization while maintaining robust training accuracy via a novel weight prediction technique, SpecTrain. Experimental results show that our proposal achieves up to 8.91x speedup compared to data parallelism on a 4-GPU platform while maintaining comparable model accuracy.