libvirt: Supporting multiple vGPU types for a single pGPU¶
Virtual GPUs in Nova was implemented in Queens but only with one supported GPU type per compute node. Now that Nested Resource Providers is a thing, this spec is discussing about how to have one supported vGPU type per physical GPU (given a GPU card can have multiple physical GPUs).
As Xen provides a specific feature where physical GPUs supporting same vGPU type are within a single pGPU group, that virt driver doesn’t need to know which exact pGPUs need to support a specific type, hence this spec only targets the libvirt driver.
Hardware vendors tell us that a physical GPU device can support multiple types. That said, Intel (to be confirmed) and NVidia vendor drivers only accept one type for all the virtual devices per graphical processing unit.
For example, NVidia GRID physical cards can accept a list of different GPU types, but the driver can only support one type per physical GPU.
Consequently, we require a way to instruct the libvirt driver which vGPU types an NVIDIA or Intel physical GPU is configured to accept.
An operator needs a way to inform the libvirt driver which vGPU types an NVIDIA or Intel physical GPU is configured to accept.
We already have
[devices]/enabled_vgpu_types that define which types the
Nova compute node can use:
enabled_vgpu_types = [str_vgpu_type_1, str_vgpu_type_2, ...]
Now we propose that libvirt will accept configuration sections that are related
to the [devices]/enabled_vgpu_types and specifies which exact pGPUs are related
to the enabled vGPU types and will have a
device_addresses option defined
List of physical PCI addresses to associate with a specific GPU type.
The particular physical GPU device address needs to be mapped to the vendor
vGPU type which that physical GPU is configured to accept. In order to
provide this mapping, there will be a CONF section with a name corresponding
to the following template: "vgpu_type_%(vgpu_type_name)s
The vGPU type to associate with the PCI devices has to be the section name
prefixed by ``vgpu_``. For example, for 'nvidia-11', you would declare
Each vGPU type also has to be declared in ``[devices]/enabled_vgpu_types``.
For example, it would be set in nova.conf:
enabled_vgpu_types = nvidia-35,nvidia-36
device_addresses = 0000:84:00.0,0000:85:00.0
device_addresses = 0000:86:00.0
In that case, the
nvidia-35 vGPU type would be supported by the physical
GPUs that are in the PCI addresses
nvidia-36 vGPU would only be supported by
If some operator messes up and provides two types for the same pGPU, an
InvalidLibvirtGPUConfig exception will be raised. If the operator forgets to
provide a type for a specific pGPU, then the first type given in
enabled_vgpu_types will be supported, like the existing situation.
If the operator fat-fingers the PCI IDs, then when creating the inventory, it
will return an exception.
As one single compute could now support multiple vGPU types, asking operators to provide host aggregates for grouping computes having the same vGPU type becomes irrelevant. Instead, we need to ask operators to amend their flavors for specific GPU capabilities if they care of such things, or Placement will just randomly pick one of the available vGPU types. For this, we propose to standardize GPU capabilities that are unfortunately very vendor specific (eg. a CUDA library version support) by having a nova.virt.vgpu_capabilities module that would translate a vendor-specific vGPU type into a set of os-traits traits. If operators want vendor-specific traits, it’s their responsibility to provide custom traits on the resource providers or ask the community to find a standard trait that would fit their needs.
Instead of creating only one inventory per pGPU, we could try to have children resource providers for a pGPU being GPU types. But then, once an instance would create a vGPU, all the other inventories but the one related to the used type would be having a total=0.
Data model impact¶
REST API impact¶
Other end user impact¶
Other deployer impact¶
Operators need to either look at the sysfs (for libvirt) for knowing the existing pGPUs and which types are supported.
Support for multiple types won’t be supported by the libvirt driver until reshaping existing single vGPU type into separate nested provider is fully done. Consequently, the recommended upgrade path is to first upgrade to Stein with a single supported vGPU type and only later modify the nova-compute service for asking for more types.
- Primary assignee:
- Other contributors:
Create the config option
Modify the libvirt virt driver code to make use of that option for creating the nested Resource Provider inventories.
Classic unittests and functional tests.
A release note will be added with a ‘feature’ section, and the Virtual GPU documentation will be modified to explain the new feature.