Many state-of-the-art ML results have been obtained by scaling up the number of parameters in existing models. However, parameters and activations for such large models often do not fit in the memory of a single accelerator device; this means that it is necessary to distribute training of large models over multiple accelerators. In this work, we propose PipeDream-2BW, a system that supports memory-efficient pipeline parallelism. PipeDream-2BW uses a novel pipelining and weight gradient coalescing strategy, combined with the double buffering of weights, to ensure high throughput, low memory footprint, and weight update semantics similar to data parallelism. In addition, PipeDream-2BW automatically partitions the model over the available hardware resources, while respecting hardware constraints such as memory capacities of accelerators and interconnect topologies. PipeDream-2BW can accelerate the training of large GPT and BERT language models by up to 20$\times$ with similar final model accuracy.
Pipeline parallelism (PP) when training neural networks enables larger models to be partitioned spatially, leading to both lower network communication and overall higher hardware utilization. Unfortunately, to preserve the statistical efficiency of sequential training, existing PP techniques sacrifice hardware efficiency by decreasing pipeline utilization or incurring extra memory costs. In this paper, we investigate to what extent these sacrifices are necessary. We devise PipeMare, a simple yet robust training method that tolerates asynchronous updates during PP execution without sacrificing …
Training deep neural networks (DNNs) in large-cluster computing environments is increasingly necessary, as networks grow in size and complexity. Local memory and processing limitations require robust data and model parallelism for crossing compute node boundaries. We propose a linear-algebraic approach to model parallelism in deep learning, which allows parallel distribution of any tensor in the DNN. Rather than rely on automatic differentiation tools, which do not universally support distributed memory parallelism models, we show that parallel data movement operations, e.g., broadcast, sum-re…
The Chinese University of Hong Kong（香港中文大学）；Huawei Technologies Co. Ltd（华为技术有限公司）
A good parallelization strategy can significantly improve the efficiency or reduce the cost for the distributed training of deep neural networks (DNNs). Recently, several methods have been proposed to find efficient parallelization strategies but they all optimize a single objective (e.g., execution time, memory consumption) and produce only one strategy. We propose FT, an efficient algorithm that searches for an optimal set of parallelization strategies to allow the trade-off among different objectives. FT can adapt to different scenarios by minimizing the memory consumption when the number o…
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 multip…
National University of Defense Technology（国防科技大学）；University of California, Los Angeles（加州大学洛杉矶分校）
We propose XPipe, an efficient asynchronous pipeline model parallelism approach for multi-GPU DNN training. XPipe is designed to use multiple GPUs to concurrently and continuously train different parts of a DNN model. To improve GPU utilization and achieve high throughput, it splits a mini-batch into a set of micro-batches. It allows the overlapping of the pipelines of multiple micro-batches, including those belonging to different mini-batches. Most importantly, the novel weight prediction strategy adopted by XPipe enables it to effectively address the weight inconsistency and staleness issues…
College of Computer and Control Engineering, Nankai University（南开大学计算机与控制工程学院）；Microsoft Research（微软研究院）
Data parallelism has emerged as a necessary technique to accelerate the training of deep neural networks (DNN). In a typical data parallelism approach, the local workers push the latest updates of all the parameters to the parameter server and pull all merged parameters back periodically. However, with the increasing size of DNN models and the large number of workers in practice, this typical data parallelism cannot achieve satisfactory training acceleration, since it usually suffers from the heavy communication cost due to transferring huge amount of information between workers and the parame…
Although deep neural networks (DNN) are able to scale with direct advances in computational power (e.g., memory and processing speed), they are not well suited to exploit the recent trends for parallel architectures. In particular, gradient descent is a sequential process and the resulting serial dependencies mean that DNN training cannot be parallelized effectively. Here, we show that a DNN may be replicated over a massive parallel architecture and used to provide a cumulative sampling of local solution space which results in rapid and robust learning. We introduce a complimentary convolution…
Microsoft Research（微软研究院）；Amazon（亚马逊）；Microsoft（微软）
Modern machine learning workloads use large models, with complex structures, that are very expensive to execute. The devices that execute complex models are becoming increasingly heterogeneous as we see a flourishing of domain-specific accelerators being offered as hardware accelerators in addition to CPUs. These trends necessitate distributing the workload across multiple devices. Recent work has shown that significant gains can be obtained with model parallelism, i.e, partitioning a neural network's computational graph onto multiple devices. In particular, this form of parallelism assumes a …
Duke University（杜克大学）；University of Southern California（南加州大学）
With the rise of artificial intelligence in recent years, Deep Neural Networks (DNNs) have been widely used in many domains. To achieve high performance and energy efficiency, hardware acceleration (especially inference) of DNNs is intensively studied both in academia and industry. However, we still face two challenges: large DNN models and datasets, which incur frequent off-chip memory accesses; and the training of DNNs, which is not well-explored in recent accelerator designs. To truly provide high throughput and energy efficient acceleration for the training of deep and large models, we ine…
Parallelization framework has become a necessity to speed up the training of deep neural networks (DNN) recently. Such framework typically employs the Model Average approach, denoted as MA-DNN, in which parallel workers conduct respective training based on their own local data while the parameters of local models are periodically communicated and averaged to obtain a global model which serves as the new start of local models. However, since DNN is a highly non-convex model, averaging parameters cannot ensure that such global model can perform better than those local models. To tackle this prob…
DNN models across many domains continue to grow in size, resulting in high resource requirements for effective training, and unpalatable (and often unaffordable) costs for organizations and research labs across scales. This paper aims to significantly reduce training costs with effective use of preemptible instances, i.e., those that can be obtained at a much cheaper price while idle, but may be preempted whenever requested by priority users. Doing so, however, requires new forms of resiliency and efficiency to cope with the possibility of frequent preemptions - a failure model that is drastic…
Deep Neural Networks (DNNs) are becoming an important tool in modern computing applications. Accelerating their training is a major challenge and techniques range from distributed algorithms to low-level circuit design. In this survey, we describe the problem from a theoretical perspective, followed by approaches for its parallelization. We present trends in DNN architectures and the resulting implications on parallelization strategies. We then review and model the different types of concurrency in DNNs: from the single operator, through parallelism in network inference and training, to distri…
Nowadays, deep neural networks (DNNs) are the core enablers for many emerging edge AI applications. Conventional approaches to training DNNs are generally implemented at central servers or cloud centers for centralized learning, which is typically time-consuming and resource-demanding due to the transmission of a large amount of data samples from the device to the remote cloud. To overcome these disadvantages, we consider accelerating the learning process of DNNs on the Mobile-Edge-Cloud Computing (MECC) paradigm. In this paper, we propose HierTrain, a hierarchical edge AI learning framework, …
We describe the neural-network training framework used in the Kaldi speech recognition toolkit, which is geared towards training DNNs with large amounts of training data using multiple GPU-equipped or multi-core machines. In order to be as hardware-agnostic as possible, we needed a way to use multiple machines without generating excessive network traffic. Our method is to average the neural network parameters periodically (typically every minute or two), and redistribute the averaged parameters to the machines for further training. Each machine sees different data. By itself, this method does …
An accelerator is a specialized integrated circuit designed to perform specific computations faster than if those were performed by CPU or GPU. A Field-Programmable DNN learning and inference accelerator (FProg-DNN) using hybrid systolic and non-systolic techniques, distributed information-control and deep pipelined structure is proposed and its microarchitecture and operation presented here. Reconfigurability attends diverse DNN designs and allows for different number of workers to be assigned to different layers as a function of the relative difference in computational load among layers. The…
To reduce training time of large-scale DNNs, scientists have started to explore parallelization strategies like data-parallelism, model-parallelism, and hybrid-parallelism. While data-parallelism has been extensively studied and developed, several problems exist in realizing model-parallelism and hybrid-parallelism efficiently. Four major problems we focus on are: 1) defining a notion of a distributed model across processes, 2) implementing forward/back-propagation across process boundaries that requires explicit communication, 3) obtaining parallel speedup on an inherently sequential task, an…
Asteroid: Resource-Efficient Hybrid Pipeline Parallelism for Collaborative DNN Training on Heterogeneous Edge Devices
School of Computer Science and Engineering, Sun Yat-sen University（中山大学计算机科学与工程学院）；Data Science and Analytics Thrust, HKUST (GZ)（香港科技大学（广州）数据科学与分析学域）；The Chinese University of Hong Kong（香港中文大学）
Supporting Safety Analysis of Image-processing DNNs through Clustering-based Approaches
SnT Centre, University of Luxembourg（卢森堡大学SnT中心）；University of Luxembourg（卢森堡大学）；University of Ottawa（渥太华大学）；Lero Centre, University of Limerick（利默里克大学Lero中心）；School of EECS, University of Ottawa（渥太华大学电子工程与计算机科学学院）
University of Cologne（科隆大学）；Center for Data and Simulation Science, University of Cologne（科隆大学数据与模拟科学中心）
Deep neural networks (DNNs) and, in particular, convolutional neural networks (CNNs) have brought significant advances in a wide range of modern computer application problems. However, the increasing availability of large amounts of datasets as well as the increasing available computational power of modern computers lead to a steady growth in the complexity and size of DNN and CNN models, respectively, and thus, to longer training times. Hence, various methods and attempts have been developed to accelerate and parallelize the training of complex network architectures. In this work, a novel CNN…
We investigate the theoretical limits of pipeline parallel learning of deep learning architectures, a distributed setup in which the computation is distributed per layer instead of per example. For smooth convex and non-convex objective functions, we provide matching lower and upper complexity bounds and show that a naive pipeline parallelization of Nesterov's accelerated gradient descent is optimal. For non-smooth convex functions, we provide a novel algorithm coined Pipeline Parallel Random Smoothing (PPRS) that is within a $d^{1/4}$ multiplicative factor of the optimal convergence rate, whe…
Foundation models are becoming the dominant deep learning technologies. Pretraining a foundation model is always time-consumed due to the large scale of both the model parameter and training dataset. Besides being computing-intensive, the training process is extremely memory-intensive and communication-intensive. These features make it necessary to apply 3D parallelism, which integrates data parallelism, pipeline model parallelism and tensor model parallelism, to achieve high training efficiency. To achieve this goal, some custom software frameworks such as Megatron-LM and DeepSpeed are develo…
University of Texas at Austin（德克萨斯大学奥斯汀分校）；Microsoft Research（微软研究院）；Microsoft（微软）；VMWare Research（威睿研究院）
Training Deep Neural Networks (DNNs) is resource-intensive and time-consuming. While prior research has explored many different ways of reducing DNN training time, the impact of input data pipeline, i.e., fetching raw data items from storage and performing data pre-processing in memory, has been relatively unexplored. This paper makes the following contributions: (1) We present the first comprehensive analysis of how the input data pipeline affects the training time of widely-used computer vision and audio Deep Neural Networks (DNNs), that typically involve complex data preprocessing. We analy…
We design and implement a ready-to-use library in PyTorch for performing micro-batch pipeline parallelism with checkpointing proposed by GPipe (Huang et al., 2019). In particular, we develop a set of design components to enable pipeline-parallel gradient computation in PyTorch's define-by-run and eager execution environment. We show that each component is necessary to fully benefit from pipeline parallelism in such environment, and demonstrate the efficiency of the library by applying it to various network architectures including AmoebaNet-D and U-Net. Our library is available at https://githu…
While distributed training significantly speeds up the training process of the deep neural network (DNN), the utilization of the cluster is relatively low due to the time-consuming data synchronizing between workers. To alleviate this problem, a novel Hierarchical Parallel SGD (HPSGD) strategy is proposed based on the observation that the data synchronization phase can be paralleled with the local training phase (i.e., Feed-forward and back-propagation). Furthermore, an improved model updating method is unitized to remedy the introduced stale gradients problem, which commits updates to the rep…
With the increased penetration and proliferation of Internet of Things (IoT) devices, there is a growing trend towards distributing the power of deep learning (DL) across edge devices rather than centralizing it in the cloud. This development enables better privacy preservation, real-time responses, and user-specific models. To deploy deep and complex models to edge devices with limited resources, model partitioning of deep neural networks (DNN) model is necessary, and has been widely studied. However, most of the existing literature only considers distributing the inference model while still …
AMDP: Asynchronous Multi-Directional Pipeline Parallelism for Large-Scale Models Training
State Key Laboratory of Blockchain and Data Security, Zhejiang University, Hangzhou, China（区块链与数据安全国家重点实验室，浙江大学，杭州，中国）；College of Computer Science and Technology, Zhejiang University, Hangzhou, China（计算机科学与技术学院，浙江大学，杭州，中国）
Orchestrating Multimodal DNN Workloads in Wireless Neural Processing
Department of Electronic and Electrical Engineering, University College London（电子与电气工程系，伦敦大学学院）；Department of Electronic Engineering, Kyung Hee University（电子工程系，庆熙大学）；School of Electrical and Electronic Engineering, the University of Sheffield（电气与电子工程学院，谢菲尔德大学）；Department of Electronics and Information Convergence Engineering, Kyung Hee University（电子与信息融合工程系，庆熙大学）
Barcelona Supercomputing Center（巴塞罗那超级计算中心）；Universitat Politècnica de Catalunya（加泰罗尼亚理工大学）；National Institute of Advanced Industrial Science and Technology（日本产业技术综合研究所）；RIKEN-CCS（理化学研究所计算科学研究中心）
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 …
BSIPipe: Bidirectional Sequence-Level Interleaved Pipeline Parallelism for Long-Sequence Large Language Model Training
State Key Laboratory of Blockchain and Data Security, Zhejiang University（浙江大学区块链与数据安全国家重点实验室）；College of Computer Science and Technology, Zhejiang University（浙江大学计算机科学与技术学院）
Model parallelism has become a necessity for training modern large-scale deep language models. In this work, we identify a new and orthogonal dimension from existing model parallel approaches: it is possible to perform pipeline parallelism within a single training sequence for Transformer-based language models thanks to its autoregressive property. This enables a more fine-grained pipeline compared with previous work. With this key idea, we design TeraPipe, a high-performance token-level pipeline parallel algorithm for synchronous model-parallel training of Transformer-based language models. W…
Distributed Deep Learning using Stochastic Gradient Staleness
Information Technology Department, Posts and Telecommunications Institute of Technology (PTIT)（信息科技部，邮电技术研究所）；Department of Mechanical and Robotics Engineering, Gwangju Institute of Science and Technology（机械与机器人工程系，全州科学技术院）
University of British Columbia（不列颠哥伦比亚大学）；Vector Institute（向量学院）；Carnegie Mellon University（卡内基梅隆大学）；University of Toronto（多伦多大学）
Data parallel training is widely used for scaling distributed deep neural network (DNN) training. However, the performance benefits are often limited by the communication-heavy parameter synchronization step. In this paper, we take advantage of the domain specific knowledge of DNN training and overlap parameter synchronization with computation in order to improve the training performance. We make two key observations: (1) the optimal data representation granularity for the communication may differ from that used by the underlying DNN model implementation and (2) different parameters can afford…
Samsung AI（三星人工智能中心）；Imperial College London（帝国理工学院）；University of Cambridge（剑桥大学）
As the use of AI-powered applications widens across multiple domains, so do increase the computational demands. Primary driver of AI technology are the deep neural networks (DNNs). When focusing either on cloud-based systems that serve multiple AI queries from different users each with their own DNN model, or on mobile robots and smartphones employing pipelines of various models or parallel DNNs for the concurrent processing of multi-modal data, the next generation of AI systems will have multi-DNN workloads at their core. Large-scale deployment of AI services and integration across mobile and…
