虚拟化环境下面向多目标优化的自适应SSD缓存系统

发布时间：2019-02-21 13:49

【摘要】：以SSD(solid state drive)为代表的新型存储介质在虚拟化环境下得到了广泛的应用,通常作为虚拟机读写缓存,起到优化磁盘I/O性能的作用.已有研究往往关注SSD缓存的容量规划,依据缓存读写命中率评价SSD缓存分配效果,未能充分考虑SSD的服务能力上限,难以适用于典型的分布式应用场景,存在虚拟机抢占SSD缓存资源,导致虚拟机中应用性能违约的可能.实现了虚拟化环境下面向多目标优化的自适应SSD缓存系统,考虑了SSD的服务能力上限.基于自适应闭环实现对虚拟机和应用状态的动态感知.动态检测局部SSD缓存抢占状态,基于聚类方法生成虚拟机的优化放置方案,依据全局SSD缓存供给能力确定虚拟机迁移顺序和时机.实验结果表明,该方法在应对典型分布式应用场景时可以有效缓解SSD缓存资源的争用,同时满足应用对虚拟机放置的需求,提升应用的性能并兼顾应用的可靠性.在Hadoop应用场景下,平均降低了25%的任务执行时间,对I/O密集型应用平均提升39%的吞吐率.在Zoo Keeper应用场景下,以不到5%的性能损失为代价,应对了虚拟化主机的单点失效带来的虚拟机宕机问题.
[Abstract]:The new storage medium, represented by SSD (solid state drive), has been widely used in virtualization environment, usually as a virtual machine read and write cache, which plays the role of optimizing disk I / O performance. Previous studies often focus on the capacity planning of SSD cache, evaluate the cache allocation effect based on cache read / write hit ratio, and fail to fully consider the upper limit of SSD service capability, so it is difficult to apply to typical distributed application scenarios. There is the possibility of virtual machine preempting SSD cache resources, resulting in default of application performance in virtual machine. An adaptive SSD cache system for multi-objective optimization in virtualized environment is implemented, and the upper limit of SSD's service capability is considered. Dynamic perception of virtual machine and application state is realized based on adaptive closed loop. The local SSD cache preemption state is dynamically detected and the optimal placement scheme of virtual machine is generated based on clustering method. The migration order and timing of virtual machine are determined according to the global SSD cache supply ability. The experimental results show that the proposed method can effectively reduce the contention of SSD cache resources, meet the requirements of virtual machine placement, improve the performance of the application and take into account the reliability of the application when dealing with typical distributed application scenarios. In the Hadoop application scenario, the average task execution time is reduced by 25%, and the average throughput for I / O intensive applications is increased by 39%. At the cost of less than 5% of the performance loss in the Zoo Keeper application scenario, the virtual machine outages caused by single point failure of the virtualized host are addressed.
【作者单位】：中国科学院软件研究所软件工程技术研发中心;计算机科学国家重点实验室(中国科学院软件研究所);中国科学院大学;
【基金】：国家重点研发计划(2016YFB1000103) 国家自然科学基金(61572480) 中国科学院青年创新促进会(2015088)~~
【分类号】：TP333

【相似文献】