基于KVM半虚拟化设备模型的客户机域间通信性能优化

发布时间：2018-06-13 20:24

本文选题：域间通信 + 内核虚拟机　；参考：《上海交通大学》2012年硕士论文

【摘要】：在以虚拟化技术支持的数据中心或分布式计算中心里,各个计算资源节点、存储资源节点之间需要进行频繁的数据传输。这些节点有些是物理机,有些是部署在物理机上的虚拟机。当若干节点部署在同一台物理机上时,称这样的节点是共生虚拟机容器域。利用共生虚拟机容器域共享物理资源这一特点,可以提升它们之间数据传输的效率,即域间通信性能的优化。通过合理的服务器配置将经常发生大量数据传输的节点部署在相同的物理机上,就可以提升整个中心系统的数据吞吐量。本文设计了一种方法来提升KVM (kernel-based virtual machine)内核虚拟机域间通信性能。该方法是基于半虚拟化设备模型和共享内存技术实现的,在虚拟机之间建立一个域间通信通道以简化和优化域间通信过程,研究内容主要可分为下面三个部分。客户机与虚拟设备之间的数据传输过程采用半虚拟化设备模型进行优化；分离的驱动以传递数据描述符取代了从客户机到宿主机的数据拷贝；同时两端利用直接通知的方法减少了I/O处理过程中的CPU模式切换操作；前端驱动可以支持客户机内核与程序直接利用网络接口进行域间通信。多个客户域之间的数据传输以一块共享的物理内存作为通道,在初始化阶段将其映射到自身的地址空间中；宿主机内核可以保证数据传输过程中不需要拷贝,可以直接对映射的地址进行读写访问；数据传输通道以特定格式的缓冲区保存并传递数据,缓冲区中除了数据外还包含了数据大小、缓冲区是否可读、缓冲区是否可写等控制信息；每个客户机拥有自己的接收频道,接收端的客户机只从共享内存通道中接收自身频道的数据；客户机发送时会解析目标地址并转化为目标客户机的频道标识符,并向该频道发送数据包；发送端的客户机与接收端的客户机通过一个轻量级的消息机制互相通知收发时数据包所在的地址以及所使用的域间通信通道的频道。除了对客户机到虚拟设备,以及虚拟设备之间的数据传输进行优化与简化外,本文还对虚拟网卡设备的内部构造进行了修改,省略不必要的数据拷贝和收发处理过程,以达到更高的域间通信效率最后根据上述方法实现了一个原型系统,并针对该系统进行了单向吞吐量、端到端响应延时的测试实验。实验证明域间通信通道模型可以提高虚拟机域间通信的性能,并发现在域间通信过程中网络接口的模拟过程带来了主要的性能开销。
[Abstract]:In the data center or distributed computing center supported by virtualization technology, each computing resource node and storage resource node need to transmit data frequently. Some of these nodes are physical machines, some are virtual machines deployed on physical machines. When several nodes are deployed on the same physical machine, such nodes are called symbiotic virtual machine container domain. By using the characteristics of sharing physical resources in the container domain of the symbiotic virtual machine, the efficiency of data transmission between them can be improved, that is, the performance optimization of inter-domain communication. The data throughput of the whole central system can be improved by deploying the nodes which often occur a large number of data transfers on the same physical machine with reasonable server configuration. In this paper, we design a method to improve inter-domain communication performance of Kernel-based virtual machine. This method is based on the paravirtualization device model and shared memory technology. An inter-domain communication channel is established between virtual machines to simplify and optimize the inter-domain communication process. The research contents can be divided into the following three parts. The process of data transmission between client and virtual device is optimized by paravirtualization device model, and the data copy from client to host is replaced by separate driver to transfer data descriptor. At the same time, the method of direct notification at both ends reduces the CPU mode switching operation in the process of I / O processing, and the front-end driver can support the communication between the client kernel and the program directly using the network interface. Data transfer between multiple client domains uses a piece of shared physical memory as a channel, which is mapped to its own address space during initialization; the host kernel ensures that no copy is required during data transmission. The mapped address can be read and write access directly; the data transfer channel saves and passes data in a buffer of a specific format, which contains the size of the data in addition to the data, whether the buffer is readable or not. Whether the buffer is writable or not, each client has its own receiving channel, and the client at the receiving end only receives the data of its own channel from the shared memory channel. When the client sends, the destination address is parsed and converted into the channel identifier of the target client, and a packet is sent to the channel; The sender client and the receiver client notify each other of the address of the packet and the channel of the inter-domain communication channel used by the sender and receiver through a lightweight messaging mechanism. In addition to optimizing and simplifying the data transmission from client to virtual device and virtual device, this paper also modifies the internal structure of virtual network card device, omits unnecessary data copy and receive process. In order to achieve higher inter-domain communication efficiency, a prototype system is implemented according to the above method, and the one-way throughput, end-to-end response delay test experiments are carried out for the system. Experiments show that the inter-domain communication channel model can improve the performance of virtual machine inter-domain communication, and it is found that the simulation process of network interface brings the main performance overhead during inter-domain communication.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP302

【相似文献】