基于应用特征的虚拟机VCPU调度优化技术研究

发布时间：2018-09-04 16:56

【摘要】：随着近年来虚拟桌面系统和云计算技术的蓬勃发展，虚拟化技术不断成熟和完善。与此同时，虚拟化环境中的应用服务也呈现多元化，并难以预测。由于虚拟机监视器对于虚拟机上运行的应用服务缺乏直观的了解，所以应用类型的不确定性使得虚拟机之间的合理资源分配变得困难。特别是，虚拟机VCPU调度算法对于I/O性能影响尤为显著。针对虚拟机监视器无法感知应用特征的问题，本文提出了基于应用特征感知的VCPU调度优化技术。通过该技术，Xen能够感知应用类型的I/O密集性特征，，为VCPU调度算法高效分配物理计算资源提供有效依据，改善I/O响应延迟。本文首先分析了Xen虚拟化平台中的关键技术，详细研究了Xen的I/O虚拟化技术和前后端分离的虚拟设备驱动模型；分析了Xen中的虚拟VCPU调度算法，论述了BOOSTING机制在调度算法处理I/O密集型应用的响应时的加速作用；针对混合型应用，提出了由于Xen虚拟机监视器与上层应用之间的“语义鸿沟”，导致了Xen无法感应到混合型应用中的I/O密集型任务，进而使得VCPU调度算法无法为I/O密集型应用的处理提供优先调度，BOOSTING机制无法发挥应有的作用，最终使得I/O响应延迟较大；为了解决上述问题，提出了基于应用特征的VCPU调度优化技术。其次，该方法通过建立基于虚拟网络后端的I/O特征感应代理，解析经过特权域转发的网络数据包，保存网络流量的历史记录。在此基础上，针对虚拟机中应用类型不可预知的特点，建立了基于置信度的I/O密集性的判定算法，综合考虑网络应用短期和长期的I/O密集性特征的变化，为测量I/O密集性提供了一种动态方法。第三，本文在建立了I/O特征测量标准的基础上，通过修改事件通知机制，为I/O请求附加了紧迫性信息。通过设计VCPU局部加速机制，改进了VCPU调度策略，使得调度算法最终能够感应混合应用中的I/O密集性特征，从而对于VCPU的调度更有针对性，有效降低了混合应用中I/O请求的响应延迟。最后，确定了I/O密集性判定算法中的相关参数，通过实验验证了在使用了基于应用特征的VCPU调度技术后，网络I/O延迟有明显下降，验证了该技术的有效性。此外，还通过实验分析了VCPU局部加速率对于网络吞吐率和CPU利用率的影响，其中，通过提高VCPU局部加速率，可以大幅提高I/O密集型应用的网络吞吐率。
[Abstract]:With the rapid development of virtual desktop system and cloud computing technology, virtualization technology is mature and perfect. At the same time, the application services in the virtualization environment are diversified and difficult to predict. Because the virtual machine monitor lacks the intuitive understanding of the application service running on the virtual machine, the uncertainty of the application type makes the reasonable resource allocation between the virtual machines difficult. In particular, the virtual machine VCPU scheduling algorithm has a significant impact on I / O performance. Aiming at the problem that virtual machine monitor can not perceive application feature, this paper proposes a VCPU scheduling optimization technology based on application feature awareness. By using this technique, Xen can perceive the I / O density of the application type, which provides an effective basis for the efficient allocation of physical computing resources by the VCPU scheduling algorithm, and improves the I / O response delay. In this paper, the key technologies of Xen virtualization platform are analyzed, the I / O virtualization technology of Xen and the virtual device driver model are studied in detail, and the virtual VCPU scheduling algorithm in Xen is analyzed. This paper discusses the accelerating effect of BOOSTING mechanism when scheduling algorithms deal with the response of I / O intensive applications, and proposes the semantic gap between Xen virtual machine monitors and upper applications for hybrid applications. As a result, Xen can not sense the I / O intensive tasks in hybrid applications, which makes the VCPU scheduling algorithm unable to provide the priority scheduling mechanism for the processing of I / O intensive applications, and finally makes the I / O response delay larger. In order to solve the above problems, a VCPU scheduling optimization technique based on application features is proposed. Secondly, by establishing the I / O feature sensing agent based on the virtual network back end, the method resolves the network packets forwarded through the privileged domain, and saves the historical records of the network traffic. On this basis, in view of the unpredictable characteristics of application types in virtual machines, an I / O denseness determination algorithm based on confidence degree is established, which takes into account the changes of I / O density characteristics of network applications in the short and long term. It provides a dynamic method for measuring I / O denseness. Thirdly, based on the standard of I / O feature measurement, the urgency information is added to the I / O request by modifying the event notification mechanism. By designing the local acceleration mechanism of VCPU, the VCPU scheduling strategy is improved, so that the scheduling algorithm can eventually sense the I / O density in hybrid applications, which is more pertinent to the scheduling of VCPU. The response latency of I / O requests in hybrid applications is reduced effectively. Finally, the relevant parameters in the algorithm of I / O denseness determination are determined. The experimental results show that the network I / O delay is significantly reduced after using the VCPU scheduling technology based on the application characteristics, and the effectiveness of this technique is verified. In addition, the effect of VCPU local acceleration on network throughput and CPU utilization is analyzed through experiments. By increasing VCPU local acceleration, the network throughput of I / O intensive applications can be greatly improved.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP302

【相似文献】