面向应用服务级目标的虚拟化资源管理

Adaptive Virtualized Resource Management for Application's SLO Guarantees

在虚拟环境中实现应用服务级目标,是当前数据中心系统管理的关键问题之一.解决该问题有两个方面的要求:一方面,在虚拟化层次和范围内,能够动态和分布式地按需调整虚拟机资源分配;另一方面,在虚拟化范围之外,能够控制由于虚拟机对非虚拟化资源的竞争所导致的性能干扰,实现虚拟机性能隔离.然而,已有工作不适用于虚拟化数据中心场景.提出一种面向应用服务级目标的虚拟化资源管理方法.首先,该方法基于反馈控制理论,通过动态调整虚拟机资源分配来实现每个应用的服务器目标;同时,还设计了一个两层结构的自适应机制,使得应用模型能够动态地捕捉虚拟机资源分配与应用性能的时变非线性关系;最后,该方法通过仲裁不同应用的资源分配请求来控制虚拟机在非虚拟化资源上的竞争干扰.实验在基于Xen的机群环境中检验了该方法在RUBiS系统和TPC-W基准上的效果.实验结果显示,该方法的应用服务级目标实现率比两种对比方法平均高29.2%,而应用服务级目标平均偏离率比它们平均低50.1%.另一方面,当RUBiS系统和TPC-W基准竞争非虚拟化的磁盘I/O资源时,该方法通过抑制TPC-W基准28.7%的处理器资源需求来优先满足RUBiS系统的磁盘I/O需求.

Virtualized resources management for service level objectives （SLOs） of applications has been one of the key problems of system management in current data centers. To solve this problem one needs to： 1） dynamically and distributed allocating resources to virtual machines （VMs） of applications on demand; 2） efficiently control interference among VMs consolidated on a single physical server due to their contention on non-virtualized resources. Many existing methods, however, are not suitable for this virtualized data center scenario. This paper presents a method for the virtualized resource management for SLOs of applications. First, based on the feedback control theory, this method can achieve SLOs of applications through dynamically resourced allocation. Second, a two-layer sell-adaptive mechanism is devised and used to dynamically capture the non-linear relationship between the performance of applications and the resources allocation. Third, this method can control the performance interference among VMs on non-virtualized resources throughvirtualized resources allocation. The study has evaluated this method on the RUBiS system and TPC-W benchmark in a Xen-based virtualized cluster. The experimental results show that the average rate of SLOs achieved by this method is 29.2% higher than ones by two existing methods. Along with the average deviation from SLOs, this method is 50.1% lower than ones of the existing methods. Furthermore, when resource contention occurs on non-virtualized disk 1/O between RUBiS and TPC-W, this method can almost entirely satisfy the disk 1/O requirement of RUBiS of high priority through restraining TPC-W requests, e.g. 28.7%, on virtualized CPU.