In distributed DNN training, parameter servers (PS) can become performance bottlenecks due to PS stragglers, caused by imbalanced parameter distribution, bandwidth contention, or computation interference. Few existing studies have investigated efficient parameter (aka load) distribution among PSs. We observe significant training inefficiency with the current parameter assignment in representative machine learning frameworks (e.g., MXNet, TensorFlow), and big potential for training acceleration with better PS load distribution. We design PSLD, a dynamic parameter server load distribution scheme, to mitigate PS straggler issues and accelerate distributed model training in the PS architecture. An exploitation-exploration method is carefully designed to scale in and out parameter servers and adjust parameter distribution among PSs on the go. We also design an elastic PS scaling module to carry out our scheme with little interruption to the training process. We implement our module on top of open-source PS architectures, including MXNet and BytePS. Testbed experiments show up to 2.86x speed-up in model training with PSLD, for different ML models under various straggler settings.

Elastic parameter server load distribution in deep learning clusters

Nowadays, large and complex deep learning (DL) models are increasingly trained in a distributed manner across multiple worker machines, in which extensive communications between workers pose serious scaling problems. In this article, we present a quantitative survey of communication optimization techniques for data parallel distributed DL. We first identify the major communication challenges and classify the existing solutions into three levels, namely the learning algorithm, the system architecture, and the network infrastructure. We present the state-of-the-art communication optimization techniques and conduct a comparative study of seven common lossless distributed DL methods on a 32-GPU cluster with 100Gb/s InfiniBand (IB). We show that the DL models with low model intensity (such as BERT and BERT-Large) are difficult to scale out even with the best available lossless algorithm over 100Gb/s IB; and the system architecture and scheduling algorithms have a critical impact on the scaling property. We conclude the article with discussions of open issues for further investigation.

A Quantitative Survey of Communication Optimizations in Distributed Deep Learning

Nowadays, large and complex deep learning (DL) models are increasingly trained in a distributed manner across multiple worker machines, in which extensive communications between workers pose serious scaling problems. In this article, we present a quantitative survey of communication optimization techniques for data parallel distributed DL. We first identify the major communication challenges and classify the existing solutions into three levels, namely the learning algorithm, the system architecture, and the network infrastructure. We present the state-of-the-art communication optimization techniques and conduct a comparative study of seven common lossless distributed DL methods on a 32-GPU cluster with 100Gbps InfiniBand (IB). We show that (1) the DL models with low model intensity (such as BERT and BERT-Large) are difficult to scale out even with the best available lossless algorithm over 100Gbps IB; (2) the system architecture and scheduling algorithms have a critical impact on the scaling property. We conclude the article with discussions on the open issues for further investigations.

The bulk synchronous parallel (BSP) is a celebrated synchronization model for general-purpose parallel computing that has successfully been employed for distributed training of machine learning models. A prevalent shortcoming of the BSP is that it requires workers to wait for the straggler at every iteration. To ameliorate this shortcoming of classic BSP, we propose ELASTICBSP a model that aims to relax its strict synchronization requirement. The proposed model offers more flexibility and adaptability during the training phase, without sacrificing on the accuracy of the trained model. We also propose an efficient method that materializes the model, named ZIPLINE. The algorithm is tunable and can effectively balance the trade-off between quality of convergence and iteration throughput, in order to accommodate different environments or applications. A thorough experimental evaluation demonstrates that our proposed ELASTICBSP model converges faster and to a higher accuracy than the classic BSP. It also achieves comparable (if not higher) accuracy than the other sensible synchronization models.

Elastic Bulk Synchronous Parallel Model for Distributed Deep Learning

Due to the additive property of most machine learning objective functions, the training can be distributed to multiple machines. Distributed machine learning is an efficient way to deal with the rapid growth of data volume at the cost of extra inter-machine communication. One common implementation is the parameter server system which contains two types of nodes: worker nodes, which are used for calculating updates, and server nodes, which are used for maintaining parameters. We observe that inefficient communication between workers and servers may slow down the system. Therefore, we propose a graph partition problem to partition data among workers and parameters among servers such that the total training time is minimized. Our problem is NP-Complete. We investigate a two-step heuristic approach that first partitions data, and then partitions parameters. We consider the trade-off between partition time and the saving in training time. Besides, we adopt a multilevel graph partition approach to fit the bipartite graph partitioning. We implement both approaches based on an open-source parameter server platform—PS-lite. Experiment results on synthetic and real-world datasets show that both approaches could significantly improve the communication efficiency up to 14 times compared with the random partition.

Minimizing Training Time of Distributed Machine Learning by Reducing Data Communication

Deep learning is a popular machine learning technique and has been applied to many real-world problems, ranging from computer vision to natural language processing. However, training a deep neural network is very time-consuming, especially on big data. It has become difficult for a single machine to train a large model over large datasets. A popular solution is to distribute and parallelize the training process across multiple machines using the parameter server framework. In this paper, we present a distributed paradigm on the parameter server framework called Dynamic Stale Synchronous Parallel (DSSP) which improves the state-of-the-art Stale Synchronous Parallel (SSP) paradigm by dynamically determining the staleness threshold at the run time. Conventionally to run distributed training in SSP, the user needs to specify a particular stalenes threshold as a hyper-parameter. However, a user does not usually know how to set the threshold and thus often finds a threshold value through trial and error, which is time-consuming. Based on workers' recent processing time, our approach DSSP adaptively adjusts the threshold per iteration at running time to reduce the waiting time of faster workers for synchronization of the globally shared parameters (the weights of the model), and consequently increases the frequency of parameters updates (increases iteration through-put), which speedups the convergence rate. We compare DSSP with other paradigms such as Bulk Synchronous Parallel (BSP), Asynchronous Parallel (ASP), and SSP by running deep neural networks (DNN) models over GPU clusters in both homogeneous and heterogeneous environments. The results show that in a heterogeneous environment where the cluster consists of mixed models of GPUs, DSSP converges to a higher accuracy much earlier than SSP and BSP and performs similarly to ASP. In a homogeneous distributed cluster, DSSP has more stable and slightly better performance than SSP and ASP, and converges much faster than BSP.

Dynamic Stale Synchronous Parallel Distributed Training for Deep Learning

Deep learning is a popular machine learning technique and has been applied to many real-world problems. However, training a deep neural network is very time-consuming, especially on big data. It has become difficult for a single machine to train a large model over large datasets. A popular solution is to distribute and parallelize the training process across multiple machines using the parameter server framework. In this paper, we present a distributed paradigm on the parameter server framework called Dynamic Stale Synchronous Parallel (DSSP) which improves the state-of-the-art Stale Synchronous Parallel (SSP) paradigm by dynamically determining the staleness threshold at the run time. Conventionally to run distributed training in SSP, the user needs to specify a particular staleness threshold as a hyper-parameter. However, a user does not usually know how to set the threshold and thus often finds a threshold value through trial and error, which is time-consuming. Based on workers' recent processing time, our approach DSSP adaptively adjusts the threshold per iteration at running time to reduce the waiting time of faster workers for synchronization of the globally shared parameters, and consequently increases the frequency of parameters updates (increases iteration throughput), which speedups the convergence rate. We compare DSSP with other paradigms such as Bulk Synchronous Parallel (BSP), Asynchronous Parallel (ASP), and SSP by running deep neural networks (DNN) models over GPU clusters in both homogeneous and heterogeneous environments. The results show that in a heterogeneous environment where the cluster consists of mixed models of GPUs, DSSP converges to a higher accuracy much earlier than SSP and BSP and performs similarly to ASP. In a homogeneous distributed cluster, DSSP has more stable and slightly better performance than SSP and ASP, and converges much faster than BSP.

Scene Classification has been addressed with numerous techniques in computer vision literature. However, with the increasing number of scene classes in datasets in the field, it has become difficult to achieve high accuracy in the context of robotics. In this paper, we implement an approach which combines traditional deep learning techniques with natural language processing methods to generate a word embedding based Scene Classification algorithm. We use the key idea that context (objects in the scene) of an image should be representative of the scene label meaning a group of objects could assist to predict the scene class. Objects present in the scene are represented by vectors and the images are re-classified based on the objects present in the scene to refine the initial classification by a Convolutional Neural Network (CNN). In our approach we address indoor Scene Classification task using a model trained with a reduced pre-processed version of the Places365 dataset and an empirical analysis is done on a real-world dataset that we built by capturing image sequences using a GoPro camera. We also report results obtained on a subset of the Places365 dataset using our approach and additionally show a deployment of our approach on a robot operating in a real-world environment.

Scene Classification in Indoor Environments for Robots using Context Based Word Embeddings.

The bulk synchronous parallel (BSP) is a celebrated synchronization model for general-purpose parallel computing that has successfully been employed for distributed training of deep learning models. A shortcoming of the BSP is that it requires workers to wait for the straggler at every iteration. Therefore, employing BSP increases the waiting time of the faster workers of a cluster and results in an overall prolonged training time. To ameliorate this shortcoming of BSP, we propose ElasticBSP, a model that aims to relax its strict synchronization requirement with an elastic synchronization by allowing delayed synchronization to minimize the waiting time. ElasticBSP offers more flexibility and adaptability during the training phase, without sacrificing the accuracy of the trained model. ElasticBSP is realized by the algorithm named ZipLine, which consists of two phases. First, it estimates for each worker the end time points of its future iterations at run time, and then a one-pass algorithm over the estimated time points of all workers is employed to fast compute an optimal future time point for synchronization. We provide theoretical results about the correctness and performance of the ZipLine algorithm. Furthermore, we propose algorithmic and implementation optimizations of ZipLine, namely ZipLineOpt and ZipLineOptBS, which reduce the time complexity of ZipLine to linearithmic time. A thorough experimental evaluation demonstrates that our proposed ElasticBSP model, materialized by the proposed optimized ZipLine variants, converges faster and to a higher accuracy than the predominant BSP. The focus of the paper is on optimizing the synchronization scheduling over a parameter server architecture. It is orthogonal to other types of optimizations, such as the learning rate optimization.

Xing Zhao

Papers

Dynamic Stale Synchronous Parallel Distributed Training for Deep Learning

Dynamic Stale Synchronous Parallel Distributed Training for Deep Learning

Scene Classification in Indoor Environments for Robots using Context Based Word Embeddings.

Elastic Bulk Synchronous Parallel Model for Distributed Deep Learning

ZipLine: an optimized algorithm for the elastic bulk synchronous parallel model