高性能网络虚拟化下的适应性中断调和系统

发布时间：2018-02-15 02:30

本文关键词： I/O虚拟化轮询中断网络　出处：《上海交通大学》2013年硕士论文　论文类型：学位论文

【摘要】：随着云计算技术和虚拟化技术的发展，，高性能网络成为云计算和数据中心中的重要构成部分。高性能网络虚拟化技术中，以单根I/O虚拟化技术为代表,已经得到了广泛应用。单根I/O虚拟化对比半虚拟化I/O性能优势明显，已经将虚拟化下的网络性能提高到接近峰值的水平。但是，现在的主流硬件虚拟化I/O系统SR-IOV依然存在着可扩展性不高的问题。研究和实验表明，SR-IOV系统的可扩展性问题和网络中断存在这很大的关系。中断虚拟化由于其多次的“陷入和模拟”开销，依旧是单根I/O虚拟化中的主要开销来源。随着网络速度的提高，产生的中断频率也就越高，开销也越来越大。本文对中断流程进行了深入分析，并且为了解决这个问题，提出了一种基于事件驱动的轮询模型来代替传统的中断模式。这个模型通过利用现有系统中存在的系统事件来驱动网络包的收发工作，从而在I/O路径中完全地取消了中断机制以及它带来的开销，同时减少了虚拟中断模式中过多上下文切换带来的缓存污染问题。根据多种多样的事件，本文提出的基于事件驱动的轮询模型可以灵活地并且可控地处理不同的网络环境。在一些系统事件不足的环境中，该系统也具有两种方法来补充不足的事件。该模型可以分别在虚拟机内部和虚拟机监控器中实现。由于本文提出的基于事件的轮询方法几乎没有额外开销，所以系统可以明显地提高单根I/O虚拟化的效率。实验证明了该系统可以获得最多59%的性能提升，以及更好的可扩展性。
[Abstract]:With the development of cloud computing technology and virtualization technology, high-performance network becomes an important component of cloud computing and data center. It's been widely used. Single root I / O virtualization has a significant performance advantage over paravirtualization, which has raised network performance in virtualization to near peak levels. However, The current mainstream hardware virtualization I / O system SR-IOV still has the problem of low scalability. The research and experiment show that the extensibility of SR-IOV system is related to the network interrupt. Interrupt virtualization is still the main source of overhead in single I / O virtualization because of its multiple "trapped and simulated" overhead. The higher the network speed, the higher the interrupt frequency. In order to solve this problem, the interrupt flow is analyzed in depth. In this paper, an event-driven polling model is proposed to replace the traditional interrupt mode, which uses the existing system events to drive the transmission and receiving of network packets. Thus, the interrupt mechanism and its overhead are completely eliminated in the I / O path, while the cache contamination caused by the excessive context switching in virtual interrupt mode is reduced. The event-driven polling model proposed in this paper can handle different network environments flexibly and controllably. The system also has two methods to supplement the insufficient events. The model can be implemented within the virtual machine and in the virtual machine monitor, respectively. Therefore, the system can significantly improve the efficiency of single-root I / O virtualization. Experiments show that the system can achieve up to 59% performance improvements and better scalability.
【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

【相似文献】