Optimization For Storage System In Xen Virtual Machine

Virtual Machine Monitor (VMM) can provide hardware abstraction that multiplexes physical resources onto multiple upper-layer guest operating systems. There has been a trend towards running services upon virtualization-based platform, such as consolidating web servers, application servers or database servers on the same physical machine. Consequently, providing strong guarantee both in fault containment and performance isolation among contenting virtual machines is of great importance. However, few research activities are conducted in performance guarantee area, especially in storage subsystem of Virtual Machine Monitors.This thesis paper presents two optimizing methods to solve fairness guarantee issues in storage system of Xen Virtual Machine Monitor: proportional CPU allocation and weighted disk service, which have not been properly addressed by previous research work. Background knowledge and related work are introduced in Chapter 1 and 2. Proportional CPU allocation framework is addressed in Chapter 3, and the weighted disk scheduling is presented in Chapter 4. Performance of these two methods is also evaluated, showing that convincing improvement of system fairness status can be achieved.The proportional CPU allocation algorithm strikes a balance between throughput and service fairness. A congestion detection method is proposed. When enough CPU resource is available, throughput takes precedence, making each domain use CPU resource up to their capacity. However, when the system is detected to be overloaded, a feedback-based adaptive control system is launched to make sure that CPU resource is proportionally designated into each domain. The dissertation implements a monitor to measure CPU consumption in Isolation Driver Domain (IDD) of Xen VMM, develops a controller to calculate proper CPU allocation, and deploys a kernel-mode actuator to enforce correct CPU utilization mode. So, graceful degradation can be achieved under CPU pressure. Maximizing IO throughput while maintaining fairness under pressure provides more satisfactory results under standard benchmarks compared with previous work.The dissertation also proposes a weighted disk scheduling algorithm, addressing the problem of lack of performance isolation in the CFQ scheduling algorithm ofXenLinux kernel. Our new scheduling algorithm, wCFQ-CRR, introduces IO weight parameter into CFQ, and replaces request dispatching number with disk service time as the metric of fairness between contenting IO request queues. Overused disk service time is compensated in each following round, providing good service time guarantee between each domain. In addition, compensating is not used when the system has only one active requesting domain, so IO throughput can be boosted under this condition. wCFQ-CRR can achieve better performance isolation than CFQ algorithm and maintain IO throughput close to CFQ as well.