Design and Implementation of Fine-Grained Scheduling Strategy for Kubernetes Heterogeneous Resources

doi:10.19678/j.issn.1000-3428.0063709

Abstract

Abstract: In a heterogeneous environment, efficient utilization of computing resources is critical for improving task efficiency and cluster utilization.As a preferred solution in container orchestration, Kubernetes lacks Graphics Processing Unit(GPU) fine-grained information in heterogeneous resource scheduling and can not satisfy user-defined requirements.Moreover, the scheduler can not perceive heterogeneous resources under the hybrid deployment of Central Processing Unit(CPU)/GPU nodes, leading to resource competition.For these reasons, a fine-grained scheduling strategy for CPU/GPU heterogeneous resources based on Kubernetes is proposed to comprehensively reflect the distribution of heterogeneous resources on nodes and hardware usage.First, GPU resource metrics are input in the scheduling algorithm, which uses the device plugin mechanism to collect detailed information regarding the GPUs on each node.Second, the filtering algorithm of the scheduler is improved based on the original filtering method of the CPU and memory, and the filtering of custom GPU information is added to screen out the nodes that satisfy the fine-grained requirements of users.Finally, for the hybrid deployment of the CPU/GPU nodes, the scoring algorithm of the scheduler is improved.The application types are dynamically sensed:the load balancing algorithm is used for CPU applications, and the least-fit algorithm is used for GPU applications.This ensures that the heterogeneous resource scheduling strategy can correctly schedule different types of applications and fully utilize the fragmented GPU node resources when CPU resources are insufficient.Experiments are conducted to verify the scheduling effect of GPU fine-grained scheduling and the scheduling under hybrid deployment of CPU/GPU nodes.The experimental results show that this strategy can effectively achieve GPU scheduling and prevent resource competition.

Key words: cloud computing, Kubernetes scheduling, heterogeneous resource, resource scheduling, CPU/GPU cluster

摘要： 在异构资源环境中高效利用计算资源是提升任务效率和集群利用率的关键。Kuberentes作为容器编排领域的首选方案，在异构资源调度场景下调度器缺少GPU细粒度信息无法满足用户自定义需求，并且CPU/GPU节点混合部署下调度器无法感知异构资源从而导致资源竞争。综合考虑异构资源在节点上的分布及其硬件状态，提出一种基于Kubernetes的CPU/GPU异构资源细粒度调度策略。利用设备插件机制收集每个节点上GPU的详细信息，并将GPU资源指标提交给调度算法。在原有CPU和内存过滤算法的基础上，增加自定义GPU信息的过滤，从而筛选出符合用户细粒度需求的节点。针对CPU/GPU节点混合部署的情况，改进调度器的打分算法，动态感知应用类型，对CPU和GPU应用分别采用负载均衡算法和最小最合适算法，保证异构资源调度策略对不同类型应用的正确调度，并且在CPU资源不足的情况下充分利用GPU节点的碎片资源。通过对GPU细粒度调度和CPU/GPU节点混合部署情况下的调度效果进行实验验证，结果表明该策略能够有效进行GPU调度并且避免资源竞争。

关键词: 云计算, Kubernetes调度, 异构资源, 资源调度, CPU/GPU集群

CLC Number:

TP391

LIU Zhibin, HUANG Qiulan, HU Qingbao, CHENG Yaodong, HU Yu, TIAN Haolai. Design and Implementation of Fine-Grained Scheduling Strategy for Kubernetes Heterogeneous Resources[J]. Computer Engineering, 2023, 49(2): 31-36,45.

刘志彬, 黄秋兰, 胡庆宝, 程耀东, 胡誉, 田浩来. Kubernetes异构资源细粒度调度策略的设计与实现[J]. 计算机工程, 2023, 49(2): 31-36,45.

/ / Recommend / Download Citations

URL: http://www.ecice06.com/EN/10.19678/j.issn.1000-3428.0063709

http://www.ecice06.com/EN/Y2023/V49/I2/31

Figures/Tables 9

References

[1] BOTTA A.Integration of cloud computing and internet of things:a survey[J].Future Generation Computer Systems, 2016, 56:684-700.
[2] SOFIA Z.Container technologies[J].Hypatia, 2000, 15(2), 181-201.
[3] GANNON D, BARGA R, SUNDARESAN N.Cloud-native applications[J].IEEE Cloud Computing, 2017, 4(5):16-21.
[4] BRENDAN B.Borg, Omega, and Kubernetes[J] Communications of the ACM, 2016, 59(5):50-57.
[5] PAHL C.Containerization and the PaaS cloud[J].IEEE Cloud Computing, 2015, 2(3):24-31.
[6] MARTINEZ G, ZEADALLY S, CHAO H C.Cloud computing service and architecture models[J].Information Sciences, 2014, 258:353-354.
[7] Kubernetes scheduler[EB/OL].[2021-12-19].https://kubernetes.io/docs/concepts/scheduling/kube-scheduler/#de ault-policies.
[8] Device plugins[EB/OL].[2021-12-19].https://kubernetes.io/docs/concepts/extend-kubernetes/compute-storage-net/device-plugins/.
[9] 董迎港.基于Kubernetes的异构资源调度的研究与实现[D].北京:北京邮电大学, 2021. DONG Y G.Research and implementation of heterogeneous resource scheduling based on Kubernetes[D].Beijing:Beijing University of Posts and Telecommunications, 2021.(in Chinese)
[10] MEDEL V, RANA O, BAÑARES J Á, et al.Modelling performance & resource management in Kubernetes[C]//Proceedings of the 9th International Conference on Utility and Cloud Computing.Washington D.C., USA:IEEE Press, 2016:257-262.
[11] MEDEL V.Characterising resource management performance in Kubernetes[J].Computers & Electrical Engineering, 2018, 68:286-297.
[12] TIWARI P K, JOSHI S.Dynamic weighted virtual machine live migration mechanism to manages load balancing in cloud computing[C]//Proceedings of IEEE International Conference on Computational Intelligence and Computing Research.Washington D.C., USA:IEEE Press, 2016:1-5.
[13] 胡程鹏, 薛涛.基于遗传算法的Kubernetes资源调度算法[J].计算机系统应用, 2021, 30(9):152-160. HU C P, XUE T.Kubernetes resource scheduling algorithm based on genetic algorithm[J].Computer Systems & Applications, 2021, 30(9):152-160.(in Chinese)
[14] ZHANG W G, MA X L, ZHANG J Z.Research on Kubernetes' resource scheduling scheme[C]//Proceedings of the 8th International Conference on Communication and Network Security.Washington D.C., USA:IEEE Press, 2018:144-148.
[15] CHANG C C, YANG S R, YEH E H, et al.A Kubernetes-based monitoring platform for dynamic cloud resource provisioning[C]//Proceedings of 2017 IEEE Global Communications Conference.Washington D.C., USA:IEEE Press, 2017:1-6.
[16] 常旭征, 焦文彬.Kubernetes资源调度算法的改进与实现[J].计算机系统应用, 2020, 29(7):256-259. CHANG X Z, JIAO W B.Improvement and implementation of Kubernetes resource scheduling algorithm[J].Computer Systems & Applications, 2020, 29(7):256-259.(in Chinese)
[17] 耿欣.深度学习云平台中GPU调度策略及任务并行化方法研究与应用[D].北京:北京邮电大学, 2020. GENG X.Research and application of GPU scheduling strategy and task parallelization method on deep learning cloud platform[D].Beijing:Beijing University of Posts and Telecommunications, 2020.(in Chinese)
[18] HARICHANE I, MAKHLOUF S A, BELALEM G.A proposal of Kubernetes scheduler using machine-learning on CPU/GPU cluster[M].Berlin, Germany:Springer, 2020.
[19] GHODSI A, ZAHARIA M, HINDMAN B, et al.Dominant resource fairness:fair allocation of multiple resource types[C]//Proceedings of the 8th USENIX Conference on Networked Systems Design and Implementation.New York, USA:ACM Press, 2011:323-336.
[20] 刘梓璇, 周建涛.负载均衡的主导资源公平分配算法[J].计算机工程与科学, 2019, 41(9):1574-1580. LIU Z X, ZHOU J T.A fair load-balanced dominant resource allocation algorithm[J].Computer Engineering & Science, 2019, 41(9):1574-1580.(in Chinese)
[21] CUI H G, ZHANG H, GANGER G R, et al.GeePS:scalable deep learning on distributed GPUs with a GPU-specialized parameter server[C]//Proceedings of the 11th European Conference on Computer Systems.New York, USA:ACM Press, 2016:1-16.
[22] DEAN J, CORRADO G S, MONGA R, et al.Large scale distributed deep networks[EB/OL].[2021-12-19].https://www.doc88.com/p-6061500891685.html.
[23] TEERAPITTAYANON S, MCDANEL B, KUNG H T.Distributed deep neural networks over the cloud, the edge and end devices[C]//Proceedings of the 37th International Conference on Distributed Computing Systems.Washington D.C., USA:IEEE Press, 2017:328-339.
[24] Schedule GPUs[EB/OL].[2021-12-19].https://kubernetes.io/docs/tasks/manage-gpus/scheduling-gpus/.
[25] NVIDIA Corporation.NVIDIA device plugin for Kubernetes[EB/OL].[2021-12-19].https://github.com/NVIDIA/k8s-device-plugin

Please choose a citation manager

Content to export