Optimizing Virtualized I/O Processing over High Speed Networks

As IT industry is shifting to cloud computing, virtualization has become the key enabling technology for server consolidation and elastic resource management in cloud infrastructures. On the other hand, increasing networking demands of cloud workloads is driving network speed rapidly migrate to 10Gbps and beyond. High speed networks require general purpose servers to provide highly efficient network processing in virtualized environment. In this paper, we start with detailed performance analysis to fully understand I/O virtualization performance on mainstream severs with 10GbE. Our experiments reveal major challenges faced by I/O virtualization over high speed networks: data movement and packet switching. Then we propose two optimizations for the scheduler inside Virtual Machine Monitor (VMM) to optimize data movement. In order to avoid cache misses on packets, we propose integrating a data movement engine considering VMM scheduling information to inject packets into right cores where corresponding guest domains are running. Lastly but not least, we also design a simplified bridge to do packet switching. Our combined optimizations significantly reduce major bottlenecks in virtualization environment and improve performance up to xx%.