Optimization For Inter-VMS Communication Performance Based On KVM With Para-Virtualized Device Model

In data centers or distributed computing centers with virtualization support, there are frequently data transmissions between every computing nodes or storage nodes. Some nodes are physical machines, the others are virtual machines which reside on physical machines. While some nodes are deployed on same physical machine, these nodes are co-resident virtual machines. The inter-virtual-machines communicating performance which refers efficiency of data transmission between co-resident virtual machines can be improved since they are sharing physical resources. If there are amounts of data transmission between two nodes, they should be deployed on same physical machine. The throughput of whole system also can be improved if the configuration is proper.This paper presents an inter-VMs communication method based on para-virtualized device model and shared-memory technique, which is designed and implemented for KVM(kernel-based virtual machine), in order to improve the performance of inter-VMs communication. A inter-VMs communication channel is built to simplify and optimize the data transmission path. The optimizing research includes three parts as following.Para-virtualized device model optimizes the device emulation and cuts the costs of data transmission between guest OS and virtual device. Split drivers decrease switches while handling I/O, and transmit data descriptors instead of data copied. The front-end driver provides network interfaces to guests'kernel and applications.All the virtual devices for inter-VMs communication map a same physical memory as inter-VMs communicating channel. They can directly operate memories for reading and writing data with shared memory mechanism of kernel. Data channel submits the network packets in formatted buffers which stores data and some controlling information including data size, read flag, write flag. Each guest has identification for using inter-VMs channel. The receiver only receives the data with its own identification from shared memory channel. The sender will transform the target MAC address to target identification and sending packets with it. There is a lightweight message mechanism between receiver and sender for noticing address of buffer and identification to the other side.Besides simplifying and optimizing the data transmission process, a modification is implemented on virtual network card structure. The data copies and deliver handling are saved to reach more effective.The unidirectional throughput, end-to-end latency and CPU utilization of a prototype system are tested. The experiment proves this virtual device model can improve inter-VMs communication performance and the major overhead occurs while device emulating.