An Oracle for Guiding Large-Scale Model/Hybrid Parallel Training of Convolutional Neural Networks

Abstract

Deep Neural Network (DNN) frameworks use distributed training to enable faster time to convergence and alleviate memory capacity limitations when training large models and/or using high dimension inputs. With the steady increase in datasets and model sizes, model/hybrid parallelism is deemed to have an important role in the future of distributed training of DNNs. We analyze the compute, communication, and memory requirements of Convolutional Neural Networks (CNNs) to understand the trade-offs between different parallelism approaches on performance and scalability. We leverage our model-driven analysis to be the basis for an oracle utility which can help in detecting the limitations and bottlenecks of different parallelism approaches at scale. We evaluate the oracle on six parallelization strategies, with four CNN models and multiple datasets (2D and 3D), on up to 1024 GPUs. The results demonstrate that the oracle has an average accuracy of about 86.74% when compared to empirical results, and as high as 97.57% for data parallelism.

The International ACM Symposium on High-Performance Parallel and Distributed Computing 2021 (HPDC'21)

References in corpus (10)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• Language Models are Few-Shot Learners
• One weird trick for parallelizing convolutional neural networks
• Breaking Band: A Breakdown of High-performance Communication
• Yet Another Accelerated SGD: ResNet-50 Training on ImageNet in 74.7 seconds
• ChainerMN: Scalable Distributed Deep Learning Framework
• torchgpipe: On-the-fly Pipeline Parallelism for Training Giant Models
• Automatic Cross-Replica Sharding of Weight Update in Data-Parallel Training
• Fully Convolutional Spatio-Temporal Models for Representation Learning in Plasma Science