Usage of new OWNER_NOVA trait¶
The spec is about the usage of new OWNER_NOVA trait.
Today in placement each resource class is typically only used by a single service. With the introduction of VGPU support in cyborg both nova and cyborg now share ownership or usage of the VGPU resource class. This creates a usage problem where different workflows are required to correctly consume the VGPU resource class based on which service created the inventory.
As both the nova and cyborg projects can report VGPU resources, we should be able to solve the shared VGPU resource class problem by each service recording that they created or “own” the resource provider, however, that is not done today.
As an operator, I would like to be able to deploy multiple services that can share the same resource class name without creating scheduling conflicts.
As an operator, I want to have a way to transition management of resources between OpenStack service using simple operations such as resizing an instance from a flavor that consumes VGPUs provided by nova to a flavor that uses VGPUs provided by cyborg.
As an end-user, I would like to be able to consume resources from nova and cyborg in a single instance without having to understand the detail of placement or scheduling.
This spec proposes adding a new OWNER_NOVA trait and pre-filter. Nova will tag every ResourceProvider it creates with a OWNER_NOVA trait, implying that inventories are provided by this service only. Nova will provide a pre-filter that actively requests its own OWNER_NOVA trait while cyborg will require the OWNER_CYBORG trait via its device profile to filter their own managed resources.
Data model impact¶
REST API impact¶
Other end user impact¶
Other deployer impact¶
That operators will see new traits for all the Resource Providers.
We need to make sure that the nova VGPU feature still works during a rolling upgrade to the nova version introducing the OWNER_NOVA trait. So to ensure that the nova compute service version will be bumped to signal when a compute service is upgraded and therefore reports the OWNER_NOVA trait on its RPs. The pre-filter filtering on OWNER_NOVA will only be enabled if the minimum compute service version indicates that every compute service is now upgraded and therefore reporting the OWNER_NOVA trait. This way during a rolling upgrade with old compute still present the pre-filter will not be enabled and the nova VGPU feature will work as today. But as soon as all the computes are upgraded the pre-filter will automatically start enforcing the OWNER_NOVA trait for all the nova VGPU feature requests.
- Primary assignee:
- Other contributors:
- Feature liaison:
Add OWNER_NOVA for os_traits project.
Tag every ResourceProvider that nova creates with a OWNER_NOVA trait.
Add pre-filter the trait for every Nova request group.
Add related unit and functional tests.
Need relate unit and functional tests.
Modify the related docs with OWNER_NOVA traits for update available resources and pre-filter scheduler.