High Performance I/O Interposition in Virtual Systems

Hypervisors implement useful features such as live migration and software-defined networking by interposing on their guest virtual machines’ I/O activity. Unfortunately, this interposition significantly reduces performance and scalability due to competition for resources between multiple guests and costly host/guest context switches. We present an efficient and scalable softwarebased I/O virtualization system that provides all of the benefits of I/O interposition while running host functionality on separate cores dedicated to serving multiple guests’ I/O. We find that two dedicated cores can interpose on the I/O activity of up to 14 I/O-intensive guests with performance that is 1.2x–3x better than the baseline, in some cases exceeding the performance of hardwarebased I/O virtualization.