|M.Sc Student||Kuperman Yossi|
|Subject||vRio: Efficient Paravirtual Remote I/O|
|Department||Department of Computer Science||Supervisor||Professor Dan Tsafrir|
|Full Thesis text|
Dedicating host cores for processing the paravirtual I/O of guest virtual machines (VMs) has been shown to greatly improve VM performance. This new virtual I/O model, however, requires the dedication of cores on each physical server, which is wasteful when there is not enough I/O activity. Likewise, the combined computational power of the dedicated per-server cores might not be enough when I/O activity is high. In the context of rack-scale computing, we propose to solve this problem by consolidating the dedicated cores spread across several servers onto one server. In this design, the processing of the paravirtual I/O of all of the VMs is done remotely, on the server that aggregates the dedicated cores. The hypervisor is therefore effectively split into two parts. We call this new model paravirtual remote I/O (vRIO), and we investigate its tradeoffs. The inherent downside of vRIO is that it prolongs I/O latency. We manage to keep it to a maximum of 18% longer than state-of-the-art paravirtualization for network I/O. The latency becomes up to 2x longer if, in addition to the dedicated core, we also migrate (and thus consolidate) high-end local block devices. The benefits are that we achieve (1) comparable throughput by using fewer dedicated cores and (2) superior throughput by using the same number of cores.