GPUs in OpenStack? It’s a long-standing question. There are many business cases for providing high-profile GPUs for every instance—namely AI, mining, and desktop. Until Queens, the only solution to expose these devices to the guests was PCI passthrough in Nova—effective, but wasteful in terms of resources.
In this session, we’ll show you how, as of Queens, you can request virtual GPUs (vGPUs) for the libvirt/KVM Nova driver. We'll share how the feature was discussed in the Nova community to integrate with the new Placement service, as well as how very different Nova drivers (Xen and libvirt/KVM) were able to cooperate to implement a major feature. We'll also discuss current feature limitations, and share the roadmap for the next releases.