On 04/12/2017 10:53 PM, 文峰sx-9149 wrote: > Will the openstack or libvirt (kvm) support NVIDIA VGPU? > I am here > <https://www.redhat.com/archives/libvir-list/2016-August/msg00939.html> to see a mail introduction libvirt kvm support VGPU. > But I do not know the current development situation of this feature. > Who can tell me about VGPU in Openstack? > Thanks. A number of things need to happen before vGPU resources are a reality in OpenStack/Nova. In order, they are: 1) Completion of the "traits framework" for resource providers [1]. This should be completed in Pike. 2) Completion of the "nested resource providers framework" [2]. This is critical because physical GPUs (or physical GPU *groups* in the case of XenServer) are child providers to the compute node resource provider and need to be tracked in a hierarchical relationship for resource accounting purposes. It is a stretch goal to get this work complete for Pike. 3) The spec for VGPU resources needs to be approved and merged [3]. This should happen today or Monday. 4) The os-traits library [4] needs to have GPU traits added to it. Jianghua from Citrix and myself are working on this. 5) The virt driver's get_inventory() methods [5] need to be reworked to account for physical GPUs (or physical GPU groups in the case of XenServer) having a set inventory of VGPU resources for each unique combination of max resolution size and other traits. 6) The flavor extra specs and image metadata need to be updated to allow an admin to configure and a user to request one or more VGPU resources from a VGPU resource provider having a set of required traits. Best, -jay [1] https://blueprints.launchpad.net/nova/+spec/resource-provider-traits [2] https://blueprints.launchpad.net/nova/+spec/nested-resource-providers [3] https://review.openstack.org/#/c/450122/ [4] https://github.com/openstack/os-traits [5] https://github.com/openstack/nova/blob/master/nova/virt/driver.py#L778