Optimizing Network Performance Of Para-virtualization Virtio Framework

Platform as a service is essentially physical resource virtualization,which is the corner stone of cloud computing.Virtualization is a resource management technology that makes underlying physical hardware transparent to users and hence high resources efficiency.Cloud computing use virtual machine clusters to integrate physical resources efficiently,enable resource sharing,and maximize resource utilization.It adds a virtual layer between the hardware layer and the application layer,which adversely affect the performance of the cloud services,such as throughput and transmission delay of data,and transactions per second(TPS).In terms of I/O virtualization,the traditional approach is to use pure software simulation to perform I/O operations.The simulation process takes a lot of CPU resources,which is not efficient.The para-virtualization approach is proposed to improve the I/O performance of virtual machines by proactively communicating between the front-end driver and the back-end driver.In KVM,a widely used virtualization platform,Virtio,a standard framework of device drivers is used to achieve para-virtualization.The framework has to switch between the guest machine and the host machine,due to the hypercalls for data conversion on behalf of the I/O requests,when processing the Virtio network requests,which takes a lot of CPU cycles and is one of the performance bottlenecks.In order to solve this problem,this paper proposed a two-end aggregation method TAM for Virtio to alleviate the adverse effects of virtual layer on performance.TAM is based on virtio-net,the network module of Virtio.At the front end of virtio-net,TAM collects requests and sends to Virtqueue,the shared channel.At the back-end of virtio-net,TAM fetches each request in Virtqueue,processes it,aggregates all the processed requests,and uses a single hypercall informing the front end asynchronously.The overhead of multiple hypercalls was reduced to a single hypercall,hence,reducing CPU overhead significantly with improved system performance.Experiments by using Netperf and Ping show that,by using TAM,the throughput of data and TPS increase by 5.76% and 4.3%,respectively,and the delay decreases by 46.2%,compared with the AHC network model,which only optimize the front-end.