A Container-Based Workflow for Distributed Training of Deep Learning Algorithms in HPC Clusters
Jose González-Abad, Álvaro López García, Valentin Y. Kozlov
备注：
This version of the article has been accepted for publication but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections. The Version of Record is available online at: https://doi.org/10.1007/s10586-022-03798-7
Deep learning has been postulated as a solution for numerous problems in different branches of science. Given the resource-intensive nature of these models, they often need to be executed on specialized hardware such graphical processing units (GPUs) in a distributed manner. In the academic field, researchers get access to this kind of resources through High Performance Computing (HPC) clusters. This kind of infrastructures make the training of these models difficult due to their multi-user nature and limited user permission. In addition, different HPC clusters may possess different peculiarities that can entangle the research cycle (e.g., libraries dependencies). In this paper we develop a workflow and methodology for the distributed training of deep learning models in HPC clusters which provides researchers with a series of novel advantages. It relies on udocker as containerization tool and on Horovod as library for the distribution of the models across multiple GPUs. udocker does not need any special permission, allowing researchers to run the entire workflow without relying on any administrator. Horovod ensures the efficient distribution of the training independently of the deep learning framework used. Additionally, due to containerization and specific features of the workflow, it provides researchers with a cluster-agnostic way of running their models. The experiments carried out show that the workflow offers good scalability in the distributed training of the models and that it easily adapts to different clusters.
Containerisation demonstrates its efficiency in application deployment in cloud computing. Containers can encapsulate complex programs with their dependencies in isolated environments, hence are being adopted in HPC clusters. HPC workload managers lack micro-services support and deeply integrated container management, as opposed to container orchestrators (e.g. Kubernetes). We introduce Torque-Operator (a plugin) which serves as a bridge between HPC workload managers and container Orchestrators.
HPC centers face increasing demand for software flexibility, and there is growing consensus that Linux containers are a promising solution. However, existing container build solutions require root privileges and cannot be used directly on HPC resources. This limitation is compounded as supercomputer diversity expands and HPC architectures become more dissimilar from commodity computing resources. Our analysis suggests this problem can best be solved with low-privilege containers. We detail relevant Linux kernel features, propose a new taxonomy of container privilege, and compare two open-sourc…
Containers improve the efficiency in application deployment and thus have been widely utilised on Cloud and lately in High Performance Computing (HPC) environments. Containers encapsulate complex programs with their dependencies in isolated environments making applications more compatible and portable. Often HPC systems have higher security levels compared to Cloud systems, which restrict users' ability to customise environments. Therefore, containers on HPC need to include a heavy package of libraries making their size relatively large. These libraries usually are specifically optimised for t…
Containerization technology offers lightweight OS-level virtualization, and enables portability, reproducibility, and flexibility by packing applications with low performance overhead and low effort to maintain and scale them. Moreover, container orchestrators (e.g., Kubernetes) are widely used in the Cloud to manage large clusters running many containerized applications. However, scheduling policies that consider the performance nuances of containerized High Performance Computing (HPC) workloads have not been well-explored yet. This paper conducts fine-grained scheduling policies for containe…
Building and deploying software on high-end computing systems is a challenging task. High performance applications have to reliably run across multiple platforms and environments, and make use of site-specific resources while resolving complicated software-stack dependencies. Containers are a type of lightweight virtualization technology that attempt to solve this problem by packaging applications and their environments into standard units of software that are: portable, easy to build and deploy, have a small footprint, and low runtime overhead. In this work we present an extension to the cont…
Software container solutions have revolutionized application development approaches by enabling lightweight platform abstractions within the so-called "containers." Several solutions are being actively developed in attempts to bring the benefits of containers to high-performance computing systems with their stringent security demands on the one hand and fundamental resource sharing requirements on the other. In this paper, we discuss the benefits and short-comings of such solutions when deployed on real HPC systems and applied to production scientific applications.We highlight use cases that a…
Variations in High Performance Computing (HPC) system software configurations mean that applications are typically configured and built for specific HPC environments. Building applications can require a significant investment of time and effort for application users and requires application users to have additional technical knowledge. Linux container technologies such as Docker and Charliecloud bring great benefits to the application development, build and deployment processes. While cloud platforms already widely support containers, HPC systems still have non-uniform support of container tec…
There is an ever-increasing need for computational power to train complex artificial intelligence (AI) & machine learning (ML) models to tackle large scientific problems. High performance computing (HPC) resources are required to efficiently compute and scale complex models across tens of thousands of compute nodes. In this paper, we discuss the issues associated with the deployment of machine learning frameworks on large scale secure HPC systems and how we successfully deployed a standard machine learning framework on a secure large scale HPC production system, to train a complex three-dimens…
HPC Containers for EBRAINS: Towards Portable Cross-Domain Software Environment
Simulation & Data Lab Neuroscience, Jülich Supercomputing Centre, Forschungszentrum Jülich（于利希研究中心）；Kirchhoff Institute for Physics, Ruprecht-Karls-Universität（海德堡大学）；ATHENA Research & Innovation Center（阿瑞塔研究与创新中心）；Computer Engineering, FernUniversität in Hagen（哈根远程大学）
Container technologies such as Docker have become a crucial component of many software industry practices especially those pertaining to reproducibility and portability. The containerization philosophy has influenced the scientific computing community, which has begun to adopt - and even develop - container technologies (such as Singularity). Leveraging containers for scientific software often poses challenges distinct from those encountered in industry, and requires different methodologies. This is especially true for HPC. With an increasing number of options for HPC in the cloud (including N…
The evolution of High-Performance Computing (HPC) platforms enables the design and execution of progressively larger and more complex workflow applications in these systems. The complexity comes not only from the number of elements that compose the workflows but also from the type of computations they perform. While traditional HPC workflows target simulations and modelling of physical phenomena, current needs require in addition data analytics (DA) and artificial intelligence (AI) tasks. However, the development of these workflows is hampered by the lack of proper programming models and envir…
Increasing popularity of the serverless computing approach has led to the emergence of new cloud infrastructures working in Container-as-a-Service (CaaS) model like AWS Fargate, Google Cloud Run, or Azure Container Instances. They introduce an innovative approach to running cloud containers where developers are freed from managing underlying resources. In this paper, we focus on evaluating capabilities of elastic containers and their usefulness for scientific computing in the scientific workflow paradigm using AWS Fargate and Google Cloud Run infrastructures. For experimental evaluation of our…
LRZ Supercomputing Centre（莱布尼茨超级计算中心）；CERN（欧洲核子研究组织）；Intel Deutschland GmbH（英特尔德国有限公司）；Intel Corp.（英特尔公司）
The increasing interest in the usage of Artificial Intelligence techniques (AI) from the research community and industry to tackle "real world" problems, requires High Performance Computing (HPC) resources to efficiently compute and scale complex algorithms across thousands of nodes. Unfortunately, typical data scientists are not familiar with the unique requirements and characteristics of HPC environments. They usually develop their applications with high-level scripting languages or frameworks such as TensorFlow and the installation process often requires connection to external systems to do…
In this paper we present a workflow management system which permits the kinds of data-driven workflows required by urgent computing, namely where new data is integrated into the workflow as a disaster progresses in order refine the predictions as time goes on. This allows the workflow to adapt to new data at runtime, a capability that most workflow management systems do not possess. The workflow management system was developed for the EU-funded VESTEC project, which aims to fuse HPC with real-time data for supporting urgent decision making. We first describe an example workflow from the VESTEC…
Science reproducibility is a cornerstone feature in scientific workflows. In most cases, this has been implemented as a way to exactly reproduce the computational steps taken to reach the final results. While these steps are often completely described, including the input parameters, datasets, and codes, the environment in which these steps are executed is only described at a higher level with endpoints and operating system name and versions. Though this may be sufficient for reproducibility in the short term, systems evolve and are replaced over time, breaking the underlying workflow reproduc…
Containers are an emerging technology that hold promise for improving productivity and code portability in scientific computing. We examine Linux container technology for the distribution of a non-trivial scientific computing software stack and its execution on a spectrum of platforms from laptop computers through to high performance computing (HPC) systems. We show on a workstation and a leadership-class HPC system that when deployed appropriately there are no performance penalties running scientific programs inside containers. For Python code run on large parallel computers, the run time is …
In this paper we would like to share our experience for transforming a parallel code for a Computational Fluid Dynamics (CFD) problem into a parallel version for the RedisDG workflow engine. This system is able to capture heterogeneous and highly dynamic environments, thanks to opportunistic scheduling strategies. We show how to move to the field of "HPC as a Service" in order to use heterogeneous platforms. We mainly explain, through the CFD use case, how to transform the parallel code and we exhibit challenges to 'unfold' the task graph dynamically in order to improve the overall performance…
There is an increasing demand to incorporate hybrid environments as part of workflows across edge, cloud, and HPC systems. In a such converging environment of cloud and HPC, containers are starting to play a more prominent role, bringing their networking infrastructure along with them. However, the current body of work shows that container overlay networks, which are often used to connect containers across physical hosts, are ill-suited for the HPC environment. They tend to impose significant overhead and noise, resulting in degraded performance and disturbance to co-processes on the same host…
Emerging non-volatile memory technologies (NVRAM) offer alternatives to hard drives that are persistent, while providing similar latencies to DRAM. Intel recently released the Optane drive, which features 3D XPoint memory technology. This device can be deployed as an SSD or as persistent memory. In this paper, we provide a performance comparison between Optane (SSD DC4800X) and NVMe (SSD DC3700) drives as block devices. We study the performance from two perspectives: 1) Benchmarking of drives using FIO workloads, and 2) Assessing the impact of using Optane over NVMe within the DataSpaces frame…
Workflows applications are becoming increasingly important to support scientific discovery. That is leading to a proliferation of workflow management systems and, thus, to a fragmented software ecosystem. Integration among existing workflow tools can improve development efficiency and, ultimately, increase the sustainability of scientific workflow software. We describe our experience with integrating RADICAL-Pilot (RP) and Parsl as a way to enable users to develop and execute workflow applications with heterogeneous tasks on heterogeneous high-performance computing resources. We describe our a…
Solving the software dependency issue under the HPC environment has always been a difficult task for both computing system administrators and application scientists. This work would like to tackle the issue by introducing the modern container technology, the Docker, to be specific. By integrating the auto-scaling feature of service discovery with the light-weight virtualization tool, the Docker, the construction of a virtual cluster on top of physical cluster hardware is attempted. Thus, through the isolation of computing environment, a remedy of software dependency of HPC environment is possi…
The HPC community is actively researching and evaluating tools to support execution of scientific applications in cloud-based environments. Among the various technologies, containers have recently gained importance as they have significantly better performance compared to full-scale virtualization, support for microservices and DevOps, and work seamlessly with workflow and orchestration tools. Docker is currently the leader in containerization technology because it offers low overhead, flexibility, portability of applications, and reproducibility. Singularity is another container solution that…
Unifying von-Neumann HPC and Neuromorphic Acceleration via the EBRAINS Research Infrastructure: A Framework for High-Performance Workflows
Simulation & Data Lab Neuroscience, Jülich Supercomputing Centre, Forschungszentrum Jülich（于利希研究中心超级计算中心神经科学与数据实验室）；Kirchhoff Institute for Physics, Ruprecht-Karls-Universität Heidelberg（海德堡大学基尔霍夫物理研究所）；ATHENA Research & Innovation Center（阿瑞娜研究与创新中心）；Computer Engineering, FernUniversität in Hagen（哈根远程大学计算机工程系）
A common workflow in science and engineering is to (i) setup and deploy large experiments with tasks comprising an application and multiple parameter values; (ii) generate intermediate results; (iii) analyze them; and (iv) reprioritize the tasks. These steps are repeated until the desired goal is achieved, which can be the evaluation/simulation of complex systems or model calibration. Due to time and cost constraints, sweeping all possible parameter values of the user application is not always feasible. Experimental Design techniques can help users reorganize submission-execution-analysis work…
This paper considers the scheduling of jobs on distributed, heterogeneous High Performance Computing (HPC) clusters. Market-based approaches are known to be efficient for allocating limited resources to those that are most prepared to pay. This context is applicable to an HPC or cloud computing scenario where the platform is overloaded. In this paper, jobs are composed of dependent tasks. Each job has a non-increasing time-value curve associated with it. Jobs are submitted to and scheduled by a market-clearing centralised auctioneer. This paper compares the performance of several policies for …
Scalable Runtime Architecture for Data-driven, Hybrid HPC and ML Workflow Applications
Rutgers, the State University of New Jersey（罗格斯大学（新泽西州立大学））；Brookhaven National Laboratory（布鲁克海文国家实验室）；Princeton Plasma Physics Laboratory（普林斯顿等离子体物理实验室）；Princeton University（普林斯顿大学）
本文扩展 RADICAL-Pilot，提出支持基于服务执行的可扩展运行时架构，以实现跨本地与远程 HPC/云资源的 HPC/ML 无缝耦合和低开销并发 ML 执行。
The recent successes and wide spread application of compute intensive machine learning and data analytics methods have been boosting the usage of the Python programming language on HPC systems. While Python provides many advantages for the users, it has not been designed with a focus on multi-user environments or parallel programming - making it quite challenging to maintain stable and secure Python workflows on a HPC system. In this paper, we analyze the key problems induced by the usage of Python on HPC clusters and sketch appropriate workarounds for efficiently maintaining multi-user Python…
Running scientific workflows on a supercomputer can be a daunting task for a scientific domain specialist. Workflow management solutions (WMS) are a standard method for reducing the complexity of application deployment on high performance computing (HPC) infrastructure. We introduce the design for a middleware system that extends and combines the functionality from existing solutions in order to create a high-level, staged user-centric operation/deployment model. This design addresses the requirements of several use cases in the life sciences, with a focus on neuroscience. In this manuscript w…
We introduce the Balsam service to manage high-throughput task scheduling and execution on supercomputing systems. Balsam allows users to populate a task database with a variety of tasks ranging from simple independent tasks to dynamic multi-task workflows. With abstractions for the local resource scheduler and MPI environment, Balsam dynamically packages tasks into ensemble jobs and manages their scheduling lifecycle. The ensembles execute in a pilot "launcher" which (i) ensures concurrent, load-balanced execution of arbitrary serial and parallel programs with heterogeneous processor requirem…
The Julia programming language has evolved into a modern alternative to fill existing gaps in scientific computing and data science applications. Julia leverages a unified and coordinated single-language and ecosystem paradigm and has a proven track record of achieving high performance without sacrificing user productivity. These aspects make Julia a viable alternative to high-performance computing's (HPC's) existing and increasingly costly many-body workflow composition strategy in which traditional HPC languages (e.g., Fortran, C, C++) are used for simulations, and higher-level languages (e.…
Rutgers University（罗格斯大学）；Brookhaven National Laboratory（布鲁克海文国家实验室）
Increasingly, scientific discovery requires sophisticated and scalable workflows. Workflows have become the ``new applications,'' wherein multi-scale computing campaigns comprise multiple and heterogeneous executable tasks. In particular, the introduction of AI/ML models into the traditional HPC workflows has been an enabler of highly accurate modeling, typically reducing computational needs compared to traditional methods. This chapter discusses various modes of integrating AI/ML models to HPC computations, resulting in diverse types of AI-coupled HPC workflows. The increasing need of couplin…
HPE HPC/AI EMEA Research Lab（慧与高性能计算/人工智能欧洲、中东和非洲研究实验室）
Artificial Intelligence (AI) applications based on Deep Neural Networks (DNN) or Deep Learning (DL) have become popular due to their success in solving problems likeimage analysis and speech recognition. Training a DNN is computationally intensive and High Performance Computing(HPC) has been a key driver in AI growth. Virtualisation and container technology have led to the convergence of cloud and HPC infrastructure. These infrastructures with diverse hardware increase the complexity of deploying and optimising AI training workloads. AI training deployments in HPC or cloud can be optimised wit…
This paper aims to create a transition path from file-based IO to streaming-based workflows for scientific applications in an HPC environment. By using the openPMP-api, traditional workflows limited by filesystem bottlenecks can be overcome and flexibly extended for in situ analysis. The openPMD-api is a library for the description of scientific data according to the Open Standard for Particle-Mesh Data (openPMD). Its approach towards recent challenges posed by hardware heterogeneity lies in the decoupling of data description in domain sciences, such as plasma physics simulations, from concret…
BiJuTy: An Interactive HPC-Aware Big Data Cluster Lifecycle Manager and Performance Assessment Utility for JupyterHub
Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI)（可扩展数据分析和人工智能中心）；Center for Interdisciplinary Digital Sciences (CIDS)（跨学科数字科学中心）；TUD Dresden University of Technology（德累斯顿工业大学）
High Performance Computing (HPC) supercomputers are expected to play an increasingly important role in HEP computing in the coming years. While HPC resources are not necessarily the optimal fit for HEP workflows, computing time at HPC centers on an opportunistic basis has already been available to the LHC experiments for some time, and it is also possible that part of the pledged computing resources will be offered as CPU time allocations at HPC centers in the future. The integration of the experiment workflows to make the most efficient use of HPC resources is therefore essential. This paper …
