基于虚拟机迁移的动态资源配置研究

发布时间：2018-01-11 16:41

本文关键词：基于虚拟机迁移的动态资源配置研究　出处：《华中科技大学》2013年博士论文　论文类型：学位论文

【摘要】：系统虚拟化技术有效解决了传统集群或数据中心存在的资源浪费现象严重、管理费用过大等问题。由于应用负载的动态性,研究负载不断变化的需求和资源供给之间的动态映射是虚拟化环境资源调度的核心问题,具有重大现实意义。实现应用需求和资源之间的动态映射主要通过虚拟机迁移方式来进行,其研究内容主要包括单虚拟机迁移优化、轻负载服务器的整合、热点消除等三方面。现有技术由于对迁移的开销、稳定性等方面的考虑不足而无法满足应用实时性和服务质量的需求,在此背景下,针对虚拟化环境中资源的有效使用和应用需求之间存在的上述三个问题设计和实现新的策略或方法显得尤为重要。虚拟机迁移是虚拟化环境动态资源配置的重要方式,分为在线迁移和非在线迁移。基于应用负载的动态变化,通过虚拟机迁移重新建立虚拟机和物理节点间的映射来实现资源的动态配置。现有虚拟机迁移主要存在传输数据量大、延迟长等缺陷,针对这个问题,一种基于页面压缩的自适应虚拟机在线迁移根据实时的网络状况,调节压缩算法的参数,保证虚拟机迁移在不同的网络环境中均能保持较好的性能。虚拟机保存/恢复是非在线迁移的一种方式,也是桌面虚拟化应用中频繁使用的操作。本文提出一种基于页面压缩的快速保存/恢复虚拟机的方法,对不同数据特征的页面采用不同的压缩方法,平衡算法的压缩率和压缩速度之间的关系,改善对网络文件系统的访问性能。实验结果显示,基于页面压缩的虚拟机保存和恢复操作分别平均减少69.4%和62.0%的时间延迟,而虚拟机内存镜像文件最多可减少83.1%的存储空间。轻负载服务器的整合是目前数据中心最主要的节能方式。现有虚拟化环境中服务器整合大多使用探索法,不能找到最优解。一种基于线性规划原理的服务器整合方法通过描述资源需求和资源提供等方面的约束条件,借助线性规划函数找到服务器整合的最优方案。同时,虚拟机迁移序列的生成方法首先识别迁移之间存在的时间依赖关系,在确保多个迁移的正确时序关系的前提下,采用并行化、基于页面压缩的迁移等方式减少资源动态配置的开销,以较少的费用将虚拟机的分布从当前状态转变到目标状态。仿真实验结果表明,服务器整合算法可有效减少激活物理机数量26.7%。热点消除是虚拟化环境中动态资源配置面临的又一个基本问题。应用负载上升后,资源竞争激烈,需要调整虚拟机和物理节点间的映射来消除热点。热点消除问题可以分成两个子问题：选择要迁移的虚拟机,从热节点上的多个可供选择的虚拟机集合中选择一个子集来进行迁移,以消除资源热点,即虚拟机选择问题；选择虚拟机后,需要为这些迁移的虚拟机确定目标节点,即虚拟机放置问题。虚拟机选择问题涉及动态资源配置的开销,而虚拟机放置问题关系动态资源配置的效果。对于虚拟机选择问题,目前大多数方法基于内存容量的粗略开销预估,选择开销较小的虚拟机进行迁移。一种基于预复制迁移的实时开销预估方法首先理论分析预复制迁移算法,确定开销的影响因素,从而建立应用特征和开销之间的关联,得出较为准确的快速迁移开销预估模型。实验结果表明,虚拟机迁移开销预估模型能快速地正确反映不同应用特征虚拟机的迁移开销的大小关系。对于虚拟机放置问题,目前大多数方法仅考虑能耗方面的因素,没有考虑虚拟机的负载变化趋势以及同一个物理节点上虚拟机之间的相互影响。一种负载变化趋势相关的虚拟机放置方法基于负载的历史信息,同时考虑虚拟机和物理机的负载变化趋势,确保虚拟机迁移的目标节点同时满足虚拟机现在和最近将来的资源请求,避免虚拟机的频繁迁移。仿真实验显示,在应用组的整个生命周期内,该方法比首次适应降序算法减少约75%的虚拟机迁移时间。
[Abstract]:Due to the dynamic nature of the application load , the dynamic mapping between the demand for changing the load and the supply of resources is the core problem of the resource scheduling of the virtualized environment . The dynamic mapping between the application requirement and the resource is mainly carried out through the virtual machine migration method . The research content of the system mainly comprises the following three aspects : single virtual machine migration optimization , light load server integration , hot spot elimination and the like . Virtual machine migration is an important way of dynamic resource allocation of virtualized environment , which is divided into online migration and non - online migration . Based on the dynamic change of application load , the virtual machine and physical node mapping are re - established through virtual machine migration to realize dynamic configuration of resources . The integration of the light load server is the most important energy saving mode in the data center . Most of the servers in the existing virtualized environment use the heuristic method to find the optimal solution . A server integration method based on the linear programming principle can find the optimal solution for the integration of the server by describing the constraint conditions such as resource demand and resource availability . A hot spot elimination is another basic problem faced by dynamic resource allocation in a virtualized environment . After the application load rises , the resources are competitive and the mapping between the virtual machine and the physical node needs to be adjusted to eliminate hot spots . A hot spot elimination problem can be divided into two sub - problems : selecting a virtual machine to be migrated , selecting a subset from a plurality of selectable virtual machine sets on the hot node to migrate , to eliminate resource hot spots , namely , selecting a virtual machine . For the choice of virtual machines , most of the methods are based on the coarse cost estimation of memory capacity and choosing the virtual machines with smaller overhead . A real - time overhead prediction method based on pre - replication migration firstly analyzes the pre - duplication migration algorithm and determines the influence factors of the overhead , and then establishes the correlation between the application characteristics and the overhead , and obtains a more accurate and fast moving cost estimation model . The experimental results show that the virtual machine migration cost estimation model can quickly and accurately reflect the size relation of the moving cost of different application characteristic virtual machines . For virtual machine placement problems , most of the methods consider only energy consumption considerations , regardless of the load variation trend of the virtual machine and the interaction between virtual machines on the same physical node . A virtual machine placement method associated with a load change trend is based on the historical information of the load , while ensuring that the target nodes of the virtual machine migrate simultaneously meet the virtual machine ' s current and future resource requests to avoid frequent migration of virtual machines . Simulation experiments show that the method reduces the virtual machine migration time by about 75 % over the first adaptive descending algorithm throughout the life cycle of the application group .

【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2013
【分类号】：TP302

【相似文献】