Network I/O fairness in virtual machines

Abstract

We present a mechanism for achieving network I/O fairness in virtual machines, by applying flexible rate limiting mechanisms directly to virtual network interfaces. Conventional approaches achieve this fairness by implementing rate limiting either in the virtual machine monitor or hypervisor, which generates considerable CPU interrupt and instruction overhead for forwarding packets. In contrast, our design pushes per-VM rate limiting as close as possible to the physical hardware themselves, effectively implementing per-virtual interface rate limiting in hardware. We show that this design reduces CPU overhead (both interrupts and instructions) by an order of magnitude. Our design can be applied either to virtual servers for cloud-based services, or to virtual routers. © 2010 ACM.