In this paper, we consider hybrid parallelism -- a paradigm that employs both Data Parallelism (DP) and Model Parallelism (MP) -- to scale distributed training of large recommendation models. We propose a compression framework called Dynamic Communication Thresholding (DCT) for communication-efficient hybrid training. DCT filters the entities to be communicated across the network through a simple hard-thresholding function, allowing only the most relevant information to pass through. For communication efficient DP, DCT compresses the parameter gradients sent to the parameter server during model synchronization. The threshold is updated only once every few thousand iterations to reduce the computational overhead of compression. For communication efficient MP, DCT incorporates a novel technique to compress the activations and gradients sent across the network during the forward and backward propagation, respectively. This is done by identifying and updating only the most relevant neurons of the neural network for each training sample in the data. We evaluate DCT on publicly available natural language processing and recommender models and datasets, as well as recommendation systems used in production at Facebook. DCT reduces communication by at least $100\times$ and $20\times$ during DP and MP, respectively. The algorithm has been deployed in production, and it improves end-to-end training time for a state-of-the-art industrial recommender model by 37\%, without any loss in performance.
We present Distributed Equivalent Substitution (DES) training, a novel distributed training framework for large-scale recommender systems with dynamic sparse features. DES introduces fully synchronous training to large-scale recommendation system for the first time by reducing communication, thus making the training of commercial recommender systems converge faster and reach better CTR. DES requires much less communication by substituting the weights-rich operators with the computationally equivalent sub-operators and aggregating partial results instead of transmitting the huge sparse weights …
Deep Neural Networks (DNNs) with sparse input features have been widely used in recommender systems in industry. These models have large memory requirements and need a huge amount of training data. The large model size usually entails a cost, in the range of millions of dollars, for storage and communication with the inference services. In this paper, we propose a hybrid hashing method to combine frequency hashing and double hashing techniques for model size reduction, without compromising performance. We evaluate the proposed models on two product surfaces. In both cases, experiment results d…
Machine learning has developed a variety of tools for learning and representing high-dimensional distributions with structure. Recent years have also seen big advances in designing multi-item mechanisms. Akin to overfitting, however, these mechanisms can be extremely sensitive to the Bayesian prior that they target, which becomes problematic when that prior is only approximately known. At the same time, even if access to the exact Bayesian prior is given, it is known that optimal or even approximately optimal multi-item mechanisms run into sample, computational, representation and communicatio…
Awerbuch et al.'s approach to distributed recommender systems (DRSs) is to have agents sample products at random while randomly querying one another for the best item they have found; we improve upon this by adding a communication network. Agents can only communicate with their immediate neighbors in the network, but neighboring agents may or may not represent users with common interests. We define two network structures: in the ``mailing-list model,'' agents representing similar users form cliques, while in the ``word-of-mouth model'' the agents are distributed randomly in a scale-free networ…
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…
Yonsei University（延世大学）；Seoul National University（首尔国立大学）；Samsung Electronics（三星电子）
In training of modern large natural language processing (NLP) models, it has become a common practice to split models using 3D parallelism to multiple GPUs. Such technique, however, suffers from a high overhead of inter-node communication. Compressing the communication is one way to mitigate the overhead by reducing the inter-node traffic volume; however, the existing compression techniques have critical limitations to be applied for NLP models with 3D parallelism in that 1) only the data parallelism traffic is targeted, and 2) the existing compression schemes already harm the model quality to…
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 impl…
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）
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…
The deep neural networks (DNNs) have been enormously successful in tasks that were hitherto in the human-only realm such as image recognition, and language translation. Owing to their success the DNNs are being explored for use in ever more sophisticated tasks. One of the ways that the DNNs are made to scale for the complex undertakings is by increasing their size -- deeper and wider networks can model well the additional complexity. Such large models are trained using model parallelism on multiple compute devices such as multi-GPUs and multi-node systems. In this paper, we develop a compiler-…
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…
ZeCO: Zero Communication Overhead Sequence Parallelism for Linear Attention
The Hong Kong Polytechnic University（香港理工大学）；TikTok；UC Santa Cruz（加州大学圣克鲁兹分校）；National University of Singapore（新加坡国立大学）；Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所）
HPC-AI Technology Inc.（HPC-AI科技公司）；University of California, Los Angeles（加利福尼亚大学洛杉矶分校）；National University of Singapore（新加坡国立大学）
Together with the improvements in state-of-the-art accuracies of various tasks, deep learning models are getting significantly larger. However, it is extremely difficult to implement these large models because limited GPU memory makes it impossible to fit large models into a single GPU or even a GPU server. Besides, it is highly necessary to reduce the training time for large models. Previous methods like Megatron-LM implemented a 1-Dimensional distributed method to use GPUs to speed up the training. However, these methods have a high communication overhead and a low scaling efficiency on larg…
We present Synkhronos, an extension to Theano for multi-GPU computations leveraging data parallelism. Our framework provides automated execution and synchronization across devices, allowing users to continue to write serial programs without risk of race conditions. The NVIDIA Collective Communication Library is used for high-bandwidth inter-GPU communication. Further enhancements to the Theano function interface include input slicing (with aggregation) and input indexing, which perform common data-parallel computation patterns efficiently. One example use case is synchronous SGD, which has rec…
The Ohio State University（俄亥俄州立大学）；Microsoft Corporation（微软公司）
In recent years, the training requirements of many state-of-the-art Deep Learning (DL) models have scaled beyond the compute and memory capabilities of a single processor, and necessitated distribution among processors. Training such massive models necessitates advanced parallelism strategies to maintain efficiency. However, such distributed DL parallelism strategies require a varied mixture of collective and point-to-point communication operations across a broad range of message sizes and scales. Examples of models using advanced parallelism strategies include Deep Learning Recommendation Mod…
Transformer achieves promising results on various tasks. However, self-attention suffers from quadratic memory requirements with respect to the sequence length. Existing work focuses on reducing time and space complexity from an algorithm perspective. In this work, we propose sequence parallelism, a memory-efficient parallelism method to help us break input sequence length limitation and train with longer sequences on GPUs efficiently. Our approach is compatible with most existing parallelisms (e.g. data parallelism, pipeline parallelism and tensor parallelism), which means our sequence parall…
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…
The recent success of deep learning applications has coincided with those widely available powerful computational resources for training sophisticated machine learning models with huge datasets. Nonetheless, training large models such as convolutional neural networks using model parallelism (as opposed to data parallelism) is challenging because the complex nature of communication between model shards makes it difficult to partition the computation efficiently across multiple machines with an acceptable trade-off. This paper presents SplitBrain, a high performance distributed deep learning fra…
Deep learning systems have become vital tools across many fields, but the increasing model sizes mean that training must be accelerated to maintain such systems' utility. Current systems like Tensorflow and MXNet focus on one specific parallelization strategy, data parallelism, which requires large training batch sizes in order to scale. We cast the problem of finding the best parallelization strategy as the problem of finding the best tiling to partition tensors with the least overall communication. We propose an algorithm that can find the optimal tiling. Our resulting parallelization soluti…
University of Wisconsin-Madison（威斯康星大学麦迪逊分校）；Carnegie Mellon University（卡内基梅隆大学）；MBZUAI（穆罕默德·本·扎耶德人工智能大学）；Petuum Inc.（Petuum公司）
Large-scale Transformer models are known for their exceptional performance in a range of tasks, but training them can be difficult due to the requirement for communication-intensive model parallelism. One way to improve training speed is to compress the message size in communication. Previous approaches have primarily focused on compressing gradients in a data parallelism setting, but compression in a model-parallel setting is an understudied area. We have discovered that model parallelism has fundamentally different characteristics than data parallelism. In this work, we present the first emp…
More than 70% of cloud computing is paid for but sits idle. A large fraction of these idle compute are cheap CPUs with few cores that are not utilized during the less busy hours. This paper aims to enable those CPU cycles to train heavyweight AI models. Our goal is against mainstream frameworks, which focus on leveraging expensive specialized ultra-high bandwidth interconnect to address the communication bottleneck in distributed neural network training. This paper presents a distributed model-parallel training framework that enables training large neural networks on small CPU clusters with lo…
