Optimization Research on Processes’ I/O Performance in Container-level Virtualization

摘要：

As representative of container-level virtualization, the core idea of OpenVZ group scheduling is fairness. System ensures CPU time is shared fairly among virtual containers and among processes belonging to the same container. Thus processes may not be scheduled immediately when awakened by I/O, affecting their responsiveness and I/O throughput. If breaking this kind of absolute fairness and permitting processes to deal with current burst I/O events by borrowing their future time temporarily, I/O performance will be improved. The approach also does not destroy system fairness in the long run. The paper analyses OpenVZ group scheduling and points out factors that affect I/O performance of processes in virtual containers. To solve these problems, we make two layers of optimization. In virtual container layer, we implement container preemption based on I/O by priority queue; in bottom layer, we still permit optimized processes to take precedence over other processes to use CPU in the case of continuous I/O, also maintaining original time mechanism and minimizing system modification. We evaluate network response time, network throughput, block I/O throughput, test results demonstrate at the scheduling cost of 2.9％ processes’ I/O performance is ameliorated remarkably.

关键词： Container-level Virtualization Group Scheduling Virtual Container Response Time Throughput Priority Queue Pair Queue

作者: Yi Zhao Taoying Liu Guanyuan Zhang Jiangning Cui

作者单位: Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China Graduate Unive Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China

会议类型: 国际会议

会议名称: Sixth International Conference on Semantics,Knowledge and Grids(第六届语义、知识与网格国际会议 SKG 2010)

会议地点: 宁波

会议语种:英文

页码: 113-120

在线出版日期: 2010-11-01（万方平台首次上网日期，不代表论文的发表时间）