基于CoreOS面向负载整合的集群调度研究  

Research on Cluster Scheduling Based on CoreOS Oriented Load Integration

作　　者：张宝婷 芮建武[1] 周鹏[1] 武延军[1] 

机构地区：[1]中国科学院软件研究所基础软件国家工程研究中心,北京100190 [2]中国科学院大学,北京100049

出　　处：《计算机系统应用》2017年第11期67-75,共9页Computer Systems & Applications

基　　金：中国科学院战略性科技先导专项课题(XDA06010600)

摘　　要：CoreOS是基于Docker的新型容器化集群服务器操作系统,发展迅速,已经得到OpenStack、Kubernetes、Salesforce、Ebay等主流云服务商的支持,云环境中负载是动态的,相应的其资源需求是动态变化的,这给集群资源高效利用带来了挑战,静态预分配峰值资源的策略带来云端资源的巨大浪费,同时空转的计算浪费大量能耗.本文提出的面向负载整合的集群调度系统(简称LICSS)实时监控集群负载分布情况,调度时使用紧凑式调度策略分配计算节点,运行时利用任务迁移技术对负载进行动态整合,实现及时收集释放空转资源降低资源能耗浪费的目的.LICSS系统设计实现了节点负载度量、任务度量、负载整合算法,并测算出节点自适应负载阈值.实验表明,LICSS系统能够根据不同时段集群负载动态变化情况对负载进行有效整合,提高了12.2%的平均资源利用率,并且基于任务整合在低负载时段触发富余节点休眠降低集群能耗.Core OS is a new containerized cluster server operating system based on Docker. It is developing rapidly and has been supported by mainstream cloud service providers such as Open Stack, kubernetes, Salesforce and Ebay. The load in the cloud environment is dynamic, therefore the resource requirement is dynamic, which poses a challenge to the efficient utilization of cluster resources. The strategy of static pre-allocation of peak resources brings a huge waste of cloud resources, and in the meantime idle calculation wastes a lot of energy consumption. In this paper, load-integrated cluster scheduling system（LICSS） is proposed to monitor the load distribution of the cluster in real time. In order to release idle resources in time to reduce energy consumption in the scheduling process, the nodes are allocated using the compact scheduling strategy and the load is dynamically integrated by the task migration technique. The LICSS system implements the node load metric, task metric, load integration algorithm, and calculates the node adaptive load threshold.Experiments show that the LICSS system can effectively integrate the load according to the dynamic changes of the cluster load in different time periods, and can improve the average resource utilization rate by 12.2%, and reduce the cluster energy consumption by triggering the dormancy of the redundant nodes during the low load period.

关 键 词：集群操作系统 负载整合 集群任务调度 负载阈值 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构] TP316[自动化与计算机技术—计算机科学与技术]