面向容器云平台的集群资源调度管理器的设计与实现

发布时间：2018-06-02 04:57

本文选题：云计算 + 容器　；参考：《浙江大学》2017年硕士论文

【摘要】：随着云计算技术的发展,以Docker为代表的容器技术为数据中心的基础架构带来了新的选择。资源调度管理对于数据中心而言有着至关重要的作用。然而,目前大多数容器云平台的资源调度系统普遍基于队列模型或二级调度框架。由于模型本身的限制,这些集群调度器在处理调度问题的全局优化上存在着一些不足。为了有效地解决这一问题,本文将集群调度问题抽象为流量网中最小费用最大流的优化问题,并将其应用在容器云平台中。在构建流量网时,我们将集群中的机器资源、工作负载等物理实体定义为网络节点,并通过对节点之间边的容量和费用进行合理赋值来将不同的调度策略实例化。特别地,部分复杂的调度策略则引入了准入控制与多轮计算。立足于容器云平台本身的特性及对资源调度问题的需求,本文主要进行了如下两个创新工作:1)在流量网中引入由租户抽象出的节点,并在此基础上提出多租户多可用域模型。该调度方法立足于容器云平台服务于多租户,横跨多个可用域的特性,通过对网络中有向弧费用的合理赋值与配置,配合准入控制与多轮计算,用以实现多租户间的差异化服务,以及作业中多实例在不同可用域内的高可用。2)对现有的协同共址模型进行功能扩展,使该模型的费用模型可以接受用户的自定义,从而具备对复杂调度策略的抽象建模能力。为了验证算法效果,本文依托Kubernetes集群进行了系列实验,涉及对不同的调度模型及集群中不同类型的工作负载的验证。实验结果证明,该算法能够良好地应用于容器云平台中。
[Abstract]:With the development of cloud computing technology, the container technology represented by Docker brings a new choice for the data center infrastructure. Resource scheduling management plays an important role in data center. However, most resource scheduling systems of container cloud platform are generally based on queue model or two-level scheduling framework. Due to the limitations of the model itself, these cluster schedulers have some shortcomings in global optimization of scheduling problems. In order to solve this problem effectively, this paper abstracts the cluster scheduling problem as the optimization problem of the minimum cost and maximum flow in the traffic network, and applies it to the container cloud platform. When constructing a traffic network, we define the physical entities such as machine resources and workload as network nodes, and instantiate different scheduling strategies by assigning reasonable capacity and cost between nodes. In particular, access control and multi-wheel computation are introduced into some complex scheduling strategies. Based on the characteristics of container cloud platform and the requirement of resource scheduling problem, this paper mainly introduces two innovative work: 1) to introduce nodes abstracted by tenants in traffic network, and on this basis, a multi-tenant multi-availability domain model is proposed. The scheduling method is based on the container cloud platform serving multi-tenant, spanning the characteristics of multiple available domains, through the reasonable assignment and configuration of the directed arc cost in the network, combined with access control and multi-round calculation. In order to realize the differentiated service between multi-tenants, and the high availability of multi-instance in different usable domains, the existing co-address model can be extended, so that the cost model of the model can accept the user's definition. Therefore, it has the ability of abstract modeling for complex scheduling policies. In order to verify the effectiveness of the algorithm, a series of experiments based on Kubernetes cluster are carried out, involving the verification of different scheduling models and different types of workloads in the cluster. Experimental results show that the algorithm can be well applied to container cloud platform.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP393.09

【相似文献】