基于Kubernetes的容器云平台资源调度策略研究

发布时间：2018-06-19 12:44

本文选题：Kubernetes + 资源调度　；参考：《电子科技大学》2017年硕士论文

【摘要】：以Docker为代表的容器技术在应用的开发、发布和部署上具有便捷性和实用性,从出现之初便受到了业界的广泛关注,由于Docker本身只注重于提供容器和镜像,因此需要一个集成的容器云管理平台高效地完成容器的编排部署、资源调度、服务发现、健康监控等任务。Kubernetes凭借其强大的容器编排能力和轻量开源的特点成为了众多容器集群调度系统的领跑者,然而Kubernetes的资源调度策略和系统自带的调度算法都较为单一,在复杂的应用场景下往往力不从心。本文在深入研究Kubernetes的核心技术后,对资源调度模块进行改进和设计,主要内容如下:1.改进了Kubernetes的资源模型。Kubernetes在进行资源调度时只考量了CPU和内存的影响,但是Pod若要正常运行还需要进行镜像下载,与持久化存储系统进行数据交互,本文在原有模型上增加了镜像下载速度和数据传输速度作为资源调度的考量因素。2.改进了Kubernetes用户绑定策略。Kubernetes的用户绑定策略较为简单,本文在此基础上设计了一种弱绑定策略,增加了用户绑定的匹配规则,支持对表达式的匹配性周期检测。3.设计抢占式调度策略。根据重启策略将Pod划分为三个优先级,当宿主机资源不足时,高优先级Pod可以抢占低优先级Pod的资源,有效的提高了高优先级Pod的运行比例。4.设计动态负载均衡调度策略。Kubernetes默认调度策略中Pod一经调度便无法迁移,本文鉴于此设计了一种基于Kubernetes的动态负载均衡改进算法,该算法的静态调度负责将待调度Pod队列的每一个Pod调度到最符合其资源描述文件的节点上;动态调度则通过监控器定期将宿主机和Pod的运行状况反馈到调度器,调度器根据系统整体负载情况作出动态调整,将负载较高的节点的一些Pod迁移到负载较低的节点上,以维持系统的整体负载均衡。
[Abstract]:The container technology represented by Docker has the convenience and practicability in the development, release and deployment of the application. From the beginning of it, the container technology has received extensive attention from the industry. Because Docker itself only pays attention to providing containers and mirrors, it needs an integrated container cloud management platform to efficiently complete the arrangement and deployment of the container, resource scheduling, Service discovery, health monitoring and other tasks.Kubernetes have become the leader of a large number of container cluster scheduling systems with its powerful container arrangement ability and light source and open source characteristics. However, the resource scheduling strategy of Kubernetes and the scheduling algorithm brought by the system are relatively simple. After studying the core technology of Kubernetes, the resource scheduling module is improved and designed. The main contents are as follows: 1. improved the resource model.Kubernetes of Kubernetes to consider only the influence of CPU and memory during the resource scheduling, but if Pod wants to run normal, it also needs to download the mirror image and make the data intersection with the persistent storage system. In the original model, this paper adds the speed of the image download and the data transmission speed as the consideration factor of the resource scheduling..2. improves the user binding strategy of the Kubernetes user binding strategy.Kubernetes. On this basis, this paper designs a weak binding strategy, adding the matching rules of the user binding, and supporting the expression. .3. design preemption scheduling strategy. According to reboot strategy, the Pod is divided into three priorities. When the host resource is insufficient, high priority Pod can seize the resources of low priority Pod, effectively improve the high priority Pod operation ratio.4. design dynamic load balancing scheduling strategy.Kubernetes default scheduling. In the strategy, Pod can not be migrated once it is dispatched. In this paper, a dynamic load balancing improvement algorithm based on Kubernetes is designed. The static scheduling of the algorithm is responsible for scheduling every Pod of the scheduled Pod queue to the node that is most consistent with its resource description file; the dynamic adjustment will regularly transport the host and Pod through the monitor. The line status is fed back to the scheduler. The scheduler makes dynamic adjustments based on the overall load condition of the system, and migrates some of the Pod with higher load to the lower load nodes to maintain the overall load balance of the system.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP393.09

【参考文献】