Support virtual GPU resources

https://blueprints.launchpad.net/nova/+spec/add-support-for-vgpu

Add support for virtual GPU (vGPU) resources.

Problem description

With some graphics virtualization solutions e.g. Intel’s GVT-g and NVIDIA GRID vGPU, a single physical Graphics Processing Unit (pGPU) can be virtualized as multiple virtual Graphics Processing Units (vGPU). Some hypervisors support to boot VMs with vGPU to accelerate graphics processing. But presently Nova can’t support vGPU.

The compute node may have one or multiple pGPUs and each pGPU could support multiple vGPUs. Some pGPUs (e.g. NVIDIA GRID K1) support several different vGPU types and each vGPU type has a fixed amount of frame buffer, number of supported display heads and maximum resolutions and are targeted at different classes of workload. Due to their different resource requirements, the maximum number of vGPUs that can be created simultaneously on a pGPU varies according to the vGPU type.

The following are examples for different vGPU types:

Example 1: vGPUs on NVIDIA GRID K1

+----------------+---------------------------------------+
| Card Type      | NVIDIA GRID K1                        |
+----------------+---------------------------------------+
| No. of pGPUs   | 4                                     |
+----------------+---------------------------------------+
| FB size (MB)   | 4096  | 2048  | 1024  | 512   | 256   |
+----------------+-------+-------+-------+-------+-------+
| Max heads      |   4   |  4    |  2    |  2    |  2    |
+----------------+-------+-------+-------+-------+-------+
| vGPU model     | K180Q | K160Q | K140Q | K120Q | K100  |
+----------------+-----------------------+---------------+
| Max Resolution |    2560x1600          |   1920x1200   |
+----------------+-----------------------+---------------+
| vGPUs per GPU  |  1    |  2    | 4     | 8     | 8     |
+----------------+-------+-------+-------+-------+-------+

Example 2: Intel GVT-g vGPUs on Intel(R) Xeon(R) CPU E3-1285 v4

+----------------+------------------------------------+
| pGPU model     | Iris Pro Graphics P6300            |
+----------------+------------------------------------+
| vGPU model     | Intel GVT-g                        |
+----------------+------------------------------------+
|Framebuffer size| 128 MB                             |
+----------------+------------------------------------+
| Max heads      | 1                                  |
+----------------+------------------------------------+
| Max Resolution | 1920x1080                          |
+----------------+------------------------------------+
| No. of vGPUs   |                                    |
|    per GPU     | 7                                  |
+----------------+------------------------------------+

In this spec, we will define a model to track vGPU resources.

Use Cases

  • As a cloud administrator, I should be able to define flavors which request an amount of vGPU resources.

  • As a cloud administrator, I should be able to specify the supported display heads number and resolutions for vGPUs defined in the flavors; end users can choose a proper flavor with the expected performance.

  • As a cloud administrator, I should be able to define flavors which request vGPUs that support some special features e.g. OpenGL to achieve hardware-accelerated rendering.

  • As an end user, I should be allowed to boot VMs which have vGPUs by using the pre-defined flavor.

Proposed change

  • Define resource tracking model for vGPU: There are both quantitative and qualitative aspects need to be tracked for vGPU resources.

    • Tracking quantitative aspects of the vGPU resource:

      • Define a new standard resource class resource-classes to track the amount of vGPUs (ResourceClass.VGPU) in the resource providers.

      • Generate the resource provider(RP) tree to track the amount of vGPUs available. The resource tracking model is as the following:

        resource provider:                  compute_node
                                        /        |          \
        resource provider:            RP_1      RP_2   ...  RP_n
                                     /           |             \
        inventory:          vGPU_inv_1       vGPU_inv_2  ...  vGPU_inv_n
        

        In virt driver (in the function of get_inventory()), it would ask the hypervisor to get the existing pGPUs, their capacity for vGPUs. With the inventory data, virt driver makes resource providers for each pGPU or each pGPU group (depend on how the pGPUs are managed by hypervisors). These resource providers will be associated as the compute_node’s children[nested-resource-providers].

        • RP for GPU: For example, libvirt will report the available vGPU number for each pGPU. In this way, if there are multiple pGPUs (same model), it can create one type of vGPUs on a pGPU and create other types of vGPUs on the remaining pGPUs.

        • RP for pGPU group: XenServer uses pGPU groups to manage pGPUs. A pGPU group is a collection of pGPUs which belong to the same model. On creating vGPU, it will search the target group for a GPU which can supply the requested vGPU. In another word, it is not possible to specify which pGPU the vGPU to be created on. So XenAPI (the virt of XenServer) should make RP for each pGPU group. And the amount of in the inventory should be total number of vGPUs which can be supplied by pGPUs belong the group.

        As described above, some pGPUs (e.g. NVIDIA GRID K1) support different sized vGPU types. The capacity for different vGPU types varies. In order to make resource tracking easier, we need to make sure the number of the vGPU is predictable. So we will add a new whitelist in nova.conf to specify the enabled vGPU types to ensure each resource provider of vGPUs only has one type of vGPUs. The whitelist is defined as the following:

        enabled_vgpu_types = [ str_vgpu_type_1, str_vgpu_type_2, ... ]
        

        Note: the str_vgpu_type_x is a string representing a vGPU type. Different hypervisors may expose the vGPU types with different strings. The virt driver should handle that properly and map the whitelist to the correct vGPUs types.

        For example, NVIDIA’s vGPU type M60-0B is exposed with the type id: “nvidia-11” in libvirt; but that’s exposed in XenServer with the type name: “GRID M60-0B”. If we want to enable this vGPU type,

        • the whitelist when libvirt is the hypervisor should be:

          enabled_vgpu_types = [ "nvidia-11" ]
          
        • the whitelist when XenServer is the hypervisor should be:

          enabled_vgpu_types = [ "GRID M60-0B" ]
          

        The vGPU resource number should be 8 (4 GPU per card * 2 vGPU per GPU). The inventory data for the resource provider for vGPUs should be as:

        {
            obj_fields.ResourceClass.vGPU: {
                "total": 8,
                "reserved": 0,
                "min_unit": 1,
                "max_unit": 1,
                "step_size": 1,
                "allocation_ratio": 1.0
            },
        }
        
    • Tracking qualitative aspects of the vGPU resources:

      • The feature of traits is targeted to support representing qualitative aspects for resources to differentiate their characteristics[os-traits].

      • GPUs also have different characteristics. We define traits for GPUs in os-traits[gpu-traits]: include maximum display heads, resolutions, features. In virt driver, it should query for the qualitative aspects of the vGPU resources; map them to the defined traits and associate these traits to the resource providers.

  • Define flavor: allow the cloud administrator to create different flavors to specify the required amount of vGPU and/or a set of required traits to meet different users’ demands.

  • Scheduler: Basing on the amount of vGPU and the required traits, the resource providers which can meet the conditions will be filtered out.

  • At spawning an instance, the virt drivers should retrieve the vGPU resource specs from the instance request specs and map them to the proper information (e.g. the GPU group in XenAPI) which is needed to create a vGPU; then create and/or associating vGPU to the instance.

Alternatives

  • It has been attempted to support vGPU by creating fake SRIOV-VF PCIs for vGPUs and then passthrough PCI devices vGPU-passthrough-PCI. But there is problem to populate the fake PCI’s address. And it can’t reflect the real situation that some vGPUs are not really PCI devices.

Data model impact

No particular data model changes needed, but it depends on the data model defined in custom-resource-classes and nested-resource-providers.

REST API impact

None

Security impact

None

Notifications impact

None

Other end user impact

None

Performance Impact

None

Other deployer impact

In order to enable the vGPU feature:

  • the operators should change the nova configure settings to enable the vGPU type for each pGPU model which will provide vGPU capabilites.

  • the operators should create new or update existing flavors to specify the amount of vGPU to be requested, and other expected traits (e.g. the dispaly resolutions, display heads, features), so that users can use different flavor to request vGPUs basing on their graphics processing demands.

  • for rolling upgrads, the operators should create or update flavors requesting vGPU after they rolled out all of their nodes into release where this spec got implemented.

Developer impact

None

Implementation

Assignee(s)

Primary assignee:

jianghuaw

Other contributors:

Work Items

  • Define standard traits into os-traits for GPUs;

  • In virt driver, add code to:

    • add whitelist for enabled vGPU types in the config file

    • query needed data for enabled vGPU types

    • generate the nested resource providers

    • generate the inventory data in resource providers

    • mapping GPU characteristics to the traits defined in os-traits

    • associate these traits to the resource providers

    • mapping traits in the boot request spec to the required metadata

    • create or/and attach vGPU to the instance basing on the metadata

Dependencies

This spec depends on the following specs to be implemented:

Testing

  • Unit tests.

Documentation Impact

Need document the configuration for vGPU.

References

History

Revisions

Release Name

Description

Queens

Introduced