Using accelerators with Cyborg

Starting from microversion 2.82, nova supports creating servers with accelerators provisioned with the Cyborg service, which provides lifecycle management for accelerators.

To launch servers with accelerators, the administrator (or an user with appropriate privileges) must do the following:

  • Create a device profile in Cyborg, which specifies what accelerator resources need to be provisioned. (See Cyborg device profiles API.)

  • Set the device profile name as an extra spec in a chosen flavor, with this syntax:

    accel:device_profile=$device_profile_name
    

    The chosen flavor may be a newly created one or an existing one.

  • Use that flavor to create a server:

    openstack server create --flavor $myflavor --image $myimage $servername
    

Nova supports only specific operations for instances with accelerators. The lists of supported and unsupported operations are as below:

  • Supported operations.

    • Creation and deletion.

    • Reboots (soft and hard).

    • Pause and unpause.

    • Stop and start.

    • Take a snapshot.

    • Backup.

    • Rescue and unrescue.

    • Rebuild.

    • Evacuate.

    • Shelve and unshelve.

  • Unsupported operations

    • Resize.

    • Suspend and resume.

    • Cold migration.

    • Live migration.

Changed in version 22.0.0(Victoria): Added support for rebuild and evacuate operations.

Changed in version 23.0.0(Wallaby): Added support for shelve and unshelve operations.

Some operations, such as lock and unlock, work as they are effectively no-ops for accelerators.

Caveats

Note

This information is correct as of the 21.0.0 Ussuri release. Where improvements have been made or issues fixed, they are noted per item.

For nested resource providers:

  • Creating servers with accelerators provisioned with the Cyborg service, if a flavor asks for resources that are provided by nested Resource Provider inventories (e.g. vGPU) and the user wants multi-create (i.e. say –max 2) then the scheduler could be returning a NoValidHosts exception even if each nested Resource Provider can support at least one specific instance, if the total wanted capacity is not supported by only one nested Resource Provider. (See bug 1874664.)

    For example, creating servers with accelerators provisioned with the Cyborg service, if two children RPs have 4 vGPU inventories each:

    • You can ask for a device profile in the flavor with 2 vGPU with –max 2.

    • But you can’t ask for a device profile in the flavor with 4 vGPU and –max 2.

Using SRIOV with Cyborg

Starting from Xena release, nova supports creating servers with SRIOV provisioned with the Cyborg service.

To launch servers with accelerators, the administrator (or an user with appropriate privileges) must do the following:

  • Create a device profile in Cyborg, which specifies what accelerator resources need to be provisioned. (See Cyborg device profiles API, Cyborg SRIOV Test Report.)

  • create a ‘accelerator-direct’ vnic type port with the device-profile name set as cyborg device profile with this syntax:

    openstack port create $port_name --network $network_name --vnic-type=accelerator-direct --device-profile $device_profile_name
    
  • create a server with that port:

    openstack server create --flavor $myflavor --image $myimage $servername --nic port-id=$port-ID
    

Nova supports only specific operations for instances with accelerators. The lists of supported and unsupported operations are as below:

  • Supported operations.

    • Creation and deletion.

    • Reboots (soft and hard).

    • Pause and unpause.

    • Stop and start.

    • Rebuild.

    • Rescue and unrescue.

    • Take a snapshot.

    • Backup.

  • Unsupported operations

    • Resize.

    • Suspend and resume.

    • Cold migration.

    • Live migration.

    • Shelve and unshelve.

    • Evacuate.

    • Attach/detach a port with device profile.