Enabling GPU passthrough

For Red Hat® OpenShift® clusters that are provisioned on hypervisors, such as VSphere, Citrix, and KVM, you can use a GPU passthrough and assign GPU resources directly to specific virtual machines (V Ms).

1. Enable GPU passthrough in the VSphere hypervisor.
   1. From the Navigator on the VSphere browser-based console, click Host > Manage.
   2. On the Hardware tab, click PCI Devices and search for the GPU devices.
   3. Click the checkbox next to all wanted GPU devices to enable passthrough and click Toggle passthrough.
      For example:

      
   Note:

   Now that GPU passthrough is enabled on the GPU devices, the next steps assign the physical GPU devices to VMs that become GPU-enabled compute nodes in the OpenShift Container Platform cluster.

2. Add or modify an existing compute node VM and assign the GPU to the compute node.
3. Edit the VM settings and expand Memory. Ensure that the Reserve all guest memory checkbox is enabled.
4. From the VM settings, select Add other device > PCI device.
5. From the New PCI device setting, select the GPU devices from the drop-down menu to assign to the VM. Click Save.
   For example:

   
6. Click VM Optionson the Edit settings page.
7. Expand Boot Options and ensure that EFI is selected from the Firmware drop-down menu.
   Note: Booting from EFI is a requirement. If booting from BIOS is configured for the VM, the GPU fails to pass through.
8. From VM Options, click Advanced. Click Edit Configuration.
9. Add the following configuration parameters to the compute node VM. Click Save.
   • pciPassthru.64bitMMIOSizeGB to 64
   • pciPassthru.use64bitMMIO to TRUE

   For example:

   
10. The VM is now assigned the specified GPU resources. Power® on the VM and run the following command to confirm that the GPU device is visible to the Guest OS:

    
    [core@worker4 ~]$ lspci | grep NVIDIA
    13:00.0 3D controller: NVIDIA Corporation GV100GL [Tesla V100 PCIe 16GB] (rev a1)
    [core@worker4 ~]$