SR-IOV Based Virtual Network Sharing Under Virtualization

Virtualization has been a key technology in Cloud Computing. It allows multipleVirtual Machines to share the same resource pool (CPU, memory and network) andincrease the hardware resource utilization efciently. Recently, researchers have beenfocusing on the fair and efcient allocation of the CPU resources between VMs butignored the network virtualization, which leads to the situation that data center lacksefcient division methods of virtual network.The network, similar to the CPU and memory, is a critical and shared resource inCloud. However, unlike the CPU and memory division strategy,virtual network band-width should be allocated according to the payment and needs of the tenants. Diferenttenants may have diferent network needs for VM.Nowadays, the most popular network charge model is the fat-rate per VM pay-ment model, which charges according to the network fow used. The virtual machinedoesn’t have the corresponding minimun network guarantee. This paper proposed areceive side network allocation mechanism under SR-IOV hardware virtualization en-vironment. The contribution of this paper is as follows:(1) First, this paper proposed three requirements that network virtualization mustmeet. We also analysis the drawbacks of the current netwowrk virtualization model incloud platform. It doesn’t provide the bandwidth guarantee of the virtual machine.(2) Second, as SR-IOV improves I/O efciency by reducing the unnecessary con-text switch and redundant data copy, we set up the KVM virtualization platform andcreate SR-IOV network virtualized virtual machines.(3) Third, this paper analyzes the receiving and sending procedure of networkpackets under NAPI. And then proposes a receive side network allocation mechanism. We use two bandwidths control methods-the interrupt Throttle rate control of eachVM and the TBF (Token Bucket Filtering) network control algorithm.(4) Last, we implement our algorithm on ixgbe and ixgbevf drivers and then eval-uate network performance of our algorithms. The result shows that the algorithm pro-portionally divides VM bandwidth according to what we specify without reducing thenetwork efciency largely.