Paravirtualization

Overview

Paravirtualization is the legacy approach to delivering virtual SSDs in which the hypervisor fully emulates the NVMe device, translating between the VM, the host OS, and the actual hardware. As framed at the 2025 OCP Global Summit, this approach has two costs: it wastes the performance of fast SSDs by consuming stack time, and it is expensive because the hypervisor overhead consumes CPU cores that could otherwise be sold — representing lost revenue for a cloud operator.

The architectural response is to remove the hypervisor from the IO path and even stop trapping the admin queue, driving toward a minimal hypervisor that reduces the attack surface from bad actors. This evolution away from paravirtualization is what the NVMe technical-proposal series (TP4165, TP4159, and the new PCIe exported NVM subsystem) is built to enable, so a virtual SSD can be migrated seamlessly without the hypervisor sitting in the data path.