Echo: Compiler-based GPU Memory Footprint Reduction for LSTM RNN Training

Abstract

The Long-Short-Term-Memory Recurrent Neural Networks (LSTM RNNs) are a popular class of machine learning models for analyzing sequential data. Their training on modern GPUs, however, is limited by the GPU memory capacity. Our profiling results of the LSTM RNN-based Neural Machine Translation (NMT) model reveal that feature maps of the attention and RNN layers form the memory bottleneck and runtime is unevenly distributed across different layers when training on GPUs. Based on these two observations, we propose to recompute the feature maps rather than stashing them persistently in the GPU memory. While the idea of feature map recomputation has been considered before, existing solutions fail to deliver satisfactory footprint reduction, as they do not address two key challenges. For each feature map recomputation to be effective and efficient, its effect on (1) the total memory footprint, and (2) the total execution time has to be carefully estimated. To this end, we propose *Echo*, a new compiler-based optimization scheme that addresses the first challenge with a practical mechanism that estimates the memory benefits of recomputation over the entire computation graph, and the second challenge by non-conservatively estimating the recomputation overhead leveraging layer specifics. *Echo* reduces the GPU memory footprint automatically and transparently without any changes required to the training source code, and is effective for models beyond LSTM RNNs. We evaluate *Echo* on numerous state-of-the-art machine learning workloads on real systems with modern GPUs and observe footprint reduction ratios of 1.89X on average and 3.13X maximum. Such reduction can be converted into faster training with a larger batch size, savings in GPU energy consumption (e.g., training with one GPU as fast as with four), and/or an increase in the maximum number of layers under the same GPU memory budget.

References in corpus (27)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• Sequence to Sequence Learning with Neural Networks
• MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
• LSTM: A Search Space Odyssey
• Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
• Deep Residual Learning for Image Recognition
• Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation
• Recurrent Neural Network Regularization
• Deep Speech 2: End-to-End Speech Recognition in English and Mandarin
• Conditional Image Generation with PixelCNN Decoders
• MXNet: A Flexible and Efficient Machine Learning Library for Heterogeneous Distributed Systems
• cuDNN: Efficient Primitives for Deep Learning
• Training Deep Nets with Sublinear Memory Cost
• Layer Normalization
• Quasi-Recurrent Neural Networks
• Neural Machine Translation in Linear Time
• Tensor Comprehensions: Framework-Agnostic High-Performance Machine Learning Abstractions
• Sockeye: A Toolkit for Neural Machine Translation
• Glow: Graph Lowering Compiler Techniques for Neural Networks
• Intel nGraph: An Intermediate Representation, Compiler, and Executor for Deep Learning
• Relay: A New IR for Machine Learning Frameworks
• Optimizing Performance of Recurrent Neural Networks on GPUs
• TBD: Benchmarking and Analyzing Deep Neural Network Training
• Memory-Efficient Backpropagation Through Time
• DLVM: A modern compiler infrastructure for deep learning systems
• Recurrent Additive Networks
• Reversible Recurrent Neural Networks

Cited by in corpus (2)

• Multi-node Bert-pretraining: Cost-efficient Approach
• AutoFL: Enabling Heterogeneity-Aware Energy Efficient Federated Learning