SR-IOV网络中的细粒度适应性中断聚集

发布时间：2018-03-11 21:34

本文选题：网络　切入点：虚拟化　出处：《上海交通大学》2014年硕士论文　论文类型：学位论文

【摘要】：在虚拟化环境中,I/O的中断处理需要通过trap-and-emulation的机制实现,然而trap-and-emulation会带来高昂的处理器开销。针对这个问题Huang等人[37]提出了基于中断聚集的AIRC算法。该算法根据系统CPU的余量,动态调整中断响应的频率,从而最大化利用CPU性能,以提高网络带宽。其中的不足表现为:一、网络带宽会受到物理设备(网卡等)限制,当网络负荷已达到实际额定值时,即便提高中断响应频率,增加处理器占用时间,也不会进一步获得网络带宽的增加;二、其中动态控制使用的主要参数需要通过实验来测定,这就影响了算法的准确度和实用价值。针对以上两方面的问题和挑战,本文开展了以下几方面的研究工作:1)基于AIRC算法,提出了改进的FAIRC算法:在FAIRC算法中,中断响应的频度是否继续提高,不仅取决于当前处理器的性能余量,还要兼顾网络带宽的余量;2)基于Xen虚拟化平台,设计并实现了面向Linux的FAIRC算法;3)基于Xen和Linux实验环境,测试了FAIRC算法在netperf测试环境下对各类网络流量的性能控制。通过在在Xen系统上运行2~8个Guest OS,及最大带宽1Gbps的网络流量情况下进行实验,结果表明该算法能保证不小于900Mbps(约90%)的带宽利用率。
[Abstract]:In a virtualized environment, the interrupt processing of I / O needs to be implemented through the mechanism of trap-and-emulation, but trap-and-emulation brings high processor overhead. In order to solve this problem, Huang et al proposed an AIRC algorithm based on interrupt aggregation, which is based on the margin of system CPU. The frequency of interrupt response is dynamically adjusted to maximize the use of CPU performance to improve the network bandwidth. Firstly, the network bandwidth is limited by physical devices (network cards, etc.), when the network load has reached the actual rating, Even if the interrupt response frequency is increased and the processor time is increased, the network bandwidth will not be further increased. Second, the main parameters used in dynamic control need to be measured experimentally. This affects the accuracy and practical value of the algorithm. In view of the above two problems and challenges, the following aspects of the research work: 1) based on the AIRC algorithm, an improved FAIRC algorithm is proposed: in the FAIRC algorithm, Whether the frequency of interrupt response continues to improve depends not only on the performance allowance of the current processor, but also on the allowance of network bandwidth. Based on Xen virtualization platform, the FAIRC algorithm for Linux is designed and implemented. It is based on Xen and Linux experimental environment. The performance control of FAIRC algorithm for all kinds of network traffic under netperf test environment is tested. Experiments are carried out under the condition of running 2 ~ 8 Guest OSs on Xen system and the maximum bandwidth of 1Gbps network traffic. The results show that this algorithm can guarantee bandwidth utilization of not less than 900 Mbpss.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP393.01

【共引文献】