基于高性能网络虚拟化的可调节信用值调度算法研究
发布时间：2018-11-06 10:39
【摘要】：虚拟化技术现在被广泛的应用于云计算来支持混合和动态负载，高速网络接口也广泛应用于各个云平台上，平台虚拟机之间的网络性能往往对云平台的整体性能产生很大的影响，因此提高高速网络I/O虚拟化的性能对促进云计算的发展有着十分重大的意义。虚拟机监管器（VMM）中的调度器在分配资源方面扮演着非常重要的角色。然而，虚拟机（VM）中的应用类型对于调度器来说是不可知的，因此I/O密集型和CPU密集型任务被同等对待，这使得虚拟机不能够充分利用像10GB以太网这样的高性能网络。 本文介绍了SR-IOV的网络解决方案，通过实验说明现在Xen默认的信用值调度器不能够有效的利用高性能网络，并找出了两个主要的原因。第一，I/O密集型虚拟机在高系统负载的情况下，得不到足够的信用值来处理I/O；第二，I/O密集型虚拟机得不到及时的调度。针对这些问题，本文提出了一种新颖的调度模型，利用两种优化来移除调度器带来的性能瓶颈。在本模型中，通过监管客户虚拟机的行为，把它们划分成了I/O密集型和CPU密集型虚拟机。I/O密集型虚拟机能够获得CPU密集型虚拟机愿意共享的额外信用值，除此之外，系统可用的信用值总量也会智能地调节来加速I/O任务的响应。 通过检测程序测试的实验结果表明在系统负载非常重的情况下，，我们的调度模型在能够在保证调度的公平性不受到过分破坏的前提下，有效地提升带宽和响应时间。
[Abstract]:Virtualization technology is now widely used in cloud computing to support hybrid and dynamic load, high-speed network interfaces are also widely used in various cloud platforms, The network performance between platform virtual machines often has a great impact on the overall performance of cloud platform, so improving the performance of high-speed network I / O virtualization is of great significance to promote the development of cloud computing. The scheduler in the hypervisor (VMM) plays a very important role in allocating resources. However, the types of applications in virtual machine (VM) are unknown to schedulers, so I / O intensive and CPU intensive tasks are treated equally, which makes virtual machines unable to make full use of high performance networks such as 10GB Ethernet. This paper introduces the network solution of SR-IOV. The experiment shows that the default credit scheduler of Xen can not utilize the high performance network effectively, and finds out two main reasons. First, I / O intensive virtual machines do not have enough credit to handle I / O under high system load; second, I / O intensive virtual machines are not scheduled in a timely manner. In order to solve these problems, a novel scheduling model is proposed, which uses two optimizations to remove the performance bottleneck caused by the scheduler. In this model, by monitoring the behavior of customer virtual machines, they are divided into I / O intensive and CPU intensive virtual machines. I / O intensive virtual machines can obtain additional credit values that CPU intensive virtual machines are willing to share. The total amount of credit available to the system will also be adjusted intelligently to accelerate the response of the I / O task. The test results show that our scheduling model can effectively improve the bandwidth and response time while ensuring that the fairness of the scheduling is not unduly damaged.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP302
本文编号：2314032
[Abstract]:Virtualization technology is now widely used in cloud computing to support hybrid and dynamic load, high-speed network interfaces are also widely used in various cloud platforms, The network performance between platform virtual machines often has a great impact on the overall performance of cloud platform, so improving the performance of high-speed network I / O virtualization is of great significance to promote the development of cloud computing. The scheduler in the hypervisor (VMM) plays a very important role in allocating resources. However, the types of applications in virtual machine (VM) are unknown to schedulers, so I / O intensive and CPU intensive tasks are treated equally, which makes virtual machines unable to make full use of high performance networks such as 10GB Ethernet. This paper introduces the network solution of SR-IOV. The experiment shows that the default credit scheduler of Xen can not utilize the high performance network effectively, and finds out two main reasons. First, I / O intensive virtual machines do not have enough credit to handle I / O under high system load; second, I / O intensive virtual machines are not scheduled in a timely manner. In order to solve these problems, a novel scheduling model is proposed, which uses two optimizations to remove the performance bottleneck caused by the scheduler. In this model, by monitoring the behavior of customer virtual machines, they are divided into I / O intensive and CPU intensive virtual machines. I / O intensive virtual machines can obtain additional credit values that CPU intensive virtual machines are willing to share. The total amount of credit available to the system will also be adjusted intelligently to accelerate the response of the I / O task. The test results show that our scheduling model can effectively improve the bandwidth and response time while ensuring that the fairness of the scheduling is not unduly damaged.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP302
【共引文献】
相关会议论文 前1条
1 董耀祖;张献涛;田坤;徐雪飞;;安腾架构硬件辅助完全虚拟化设计与实现[A];2007系统仿真技术及其应用学术会议论文集[C];2007年
相关硕士学位论文 前5条
1 孙世昶;虚拟化技术研究及PCI设备直接访问的实现[D];大连理工大学;2006年
2 姚远;Pcanel/V2-基于Intel VT-x技术的VMM架构[D];浙江大学;2007年
3 杨楠;Lyn：新型轻量级虚拟化技术[D];浙江大学;2007年
4 宋曦宁;基于Xen的虚拟化办公应用环境研究[D];中国科学院研究生院（沈阳计算技术研究所）;2007年
5 朱鸿伟;虚拟化安全关键技术研究[D];浙江大学;2008年
本文编号：2314032
本文链接：https://www.wllwen.com/kejilunwen/jisuanjikexuelunwen/2314032.html
