Optimizing Network I/O Virtualization for Scale-out Processor

Recently, the idea of using ARM processors in servers appeals to be attractive since it may provide another choice to decrease power consumption in cloud platform. Using virtualization, Cloud computing decouples system resources from physical machines. Delivering interrupts among virtual machines efficiently is a key factor to achieve high performance in the cloud platform. However, we observed that the network I/O performance would dramatically fall due to the inefficient virtual interrupt handling on ARM platform when the virtual machines handle concurrent requests. In this paper, we first analyze and compare interrupt controllers and message-based interrupt delivery in both x86 and ARM architecture. With the help of hardware-assist virtual interrupt controller, we propose an optimization method that enables KVM ARM to delivery network packets through multiple virtual I/O channels by using virtual MSI notification mechanism. The experiment shows that it can improve the interrupt handling in virtual machines on ARM platform as well as the scalability of packets processing on multicore virtual machine. The throughput of concurrent requests handling can be improved by about 50% when the number of cores for the virtual machine is set to 12.