面向多处理器虚拟机的动态NUMA方法  

Dynamic NUMA on Multi-processor Hypervisor

作　　者：施继成[1,2] 陈海波[3] 臧斌宇[3] 

机构地区：[1]复旦大学软件学院,上海201203 [2]复旦大学上海市数据科学重点实验室,上海201203 [3]上海交通大学软件学院并行与分布式系统研究所,上海200240

出　　处：《小型微型计算机系统》2015年第4期677-682,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61003002)资助

摘　　要：NUMA内存技术在增加了计算机系统内存带宽的同时,其跨节点访存延迟高的特点也增加了虚拟化平台的内存管理难度.现有的多核虚拟机NUMA优化方法大多为静态方法,即无需动态调节虚拟机状态.这些方法大都限制了虚拟化监控器调度虚拟CPU的灵活性,降低了虚拟化平台负载均衡的能力,而影响虚拟机的运行效率.针对该问题,提出动态NUMA方法——在告知虚拟机NUMA拓扑结构的情况下,该方法通过调节虚拟机内操作系统NUMA相关数据结构方法,达到配合虚拟机监控器任务调度器的目的,实现虚拟机运行效率的提升.相较于静态方法,本文提出调节虚拟机NUMA数据结构,以适应虚拟机监控器调度器的动态方法.此方法提高了虚拟CPU的调度灵活性,提高了虚拟化平台负载均衡的能力,从而保证了多处理器虚拟机的运行效率.NUMA architecture brings not orfly larger memory bandwidth, but also higher latency of cross-node memory access, that makes memory management harder in virtualization. Nowadays most methods to optimize NUMA performance of multi-processor VM is static, which means not changing VM＇s states when it＇s running. These methods limit the flexibility of VMM scheduler, reduce the possibility of VM load-balance and finally affect the VM efficiency. To solve this problem, we propose dynamic NUMA, a method to adjust NUMA related data structure in VM according to VMM＇s scheduler. This method needs VM aware of NUMA topology. It＇s the first time of proposing a dynamic method to optimize NUMA performance in virtualization environment. Compared to the static meth- ods, our method improves the flexibility of VMM scheduler, makes VM load-balance easier and improves the performance of multi- processor VM in the end.

关 键 词：虚拟机监控器 多处理器 虚拟化 NUMA 

分 类 号：TP333[自动化与计算机技术—计算机系统结构]