基于可控中断计算和高效全局资源亲和性的网络I/O虚拟化方法研究

High-performance network cards are important network devices for processing numerous amounts of data in cloud data centers in the big data era. Given the low resource utilization rate of network cards, research on the I/O virtualization for high-performance network cards has been carried out. However, the performance of network I/O virtualization is lagging behind compared with that of CPU and memory virtualization, which leads to high I/O response latency and low throughput for complex and varying network data computation. To improve the performance of network I/O virtualization, this study focuses on interrupt handling methods and global resource affinity methods for network I/O virtualization. We will propose an efficient global resource (including virtual CPU, RAM, high-performance network cards) affinity method for network I/O virtualization，and also present the technology of interrupt holder (i.e., virtual CPU that holds the interrupt) detection as well as the delay control scheduling to guarantee the robustness of interrupt remapping. We will study the resource scheduling of fine-grained application perceptive virtual machines. By considering the QoS requests of users and the load of systems, we apply real-time perception and fine-grained multi-classification to the virtual machines, moreover, we allocate appropriate physical resources. Based on the method stated above, we will aim to design a controllable and efficient fine-grained application perceptive network I/O virtualization method to provide theoretical support and a technical breakthrough for the network I/O virtualization technology.

高性能网卡是大数据时代云数据中心处理海量数据的重要网络设备，由于网卡的资源利用率低，对高性能网卡的网络I/O虚拟化研究已经开展。但是网络I/O虚拟化的性能相对于CPU、内存虚拟化仍较为滞后，对复杂多变的网络数据计算，表现出高I/O响应延时和低网络吞吐量。为了提升网络I/O虚拟化的性能，本项目以网络I/O虚拟化的中断处理方法和全局资源间亲和性方法作为研究重点，拟提出高效的全局资源（虚拟CPU、内存、高性能网卡）亲和性的网络I/O虚拟化方法，并拟提出中断持有者（持有中断的虚拟CPU）检测、延时可控调度技术，保证中断重映射的鲁棒性。拟研究细粒度应用感知的虚拟机资源调度，从用户QoS需求与系统负载两个维度，对计算网络数据的虚拟机进行实时感知、细粒度多分类，并分配合理的物理资源。基于上述方法，拟设计一种细粒度应用感知的可控、高效的网络I/O虚拟化方法，为网络I/O虚拟化技术提供理论支持和技术突破。

高性能网卡是大数据时代云数据中心处理海量数据的重要网络设备，由于网卡的资源利用率低，对高性能网卡的网络I/O虚拟化研究已经开展。但是网络I/O虚拟化的性能相对于CPU、内存虚拟化仍较为滞后，对复杂多变的网络数据计算，表现出高I/O响应延时和低网络吞吐量。为了提升网络I/O虚拟化的性能，本项目以网络I/O虚拟化的中断处理方法和全局资源间亲和性方法作为研究重点，提出高效的全局资源（虚拟CPU、内存、高性能网卡）亲和性的网络I/O虚拟化方法，并提出中断持有者（持有中断的虚拟CPU）检测、延时可控调度技术，保证中断重映射的鲁棒性。本文所研究细粒度应用感知的虚拟机资源调度，从用户QoS需求与系统负载两个维度，对计算网络数据的虚拟机进行实时感知、细粒度多分类，并分配合理的物理资源。基于上述方法，本文设计一种细粒度应用感知的可控、高效的网络I/O虚拟化方法，为网络I/O虚拟化技术提供理论支持和技术突破。