Headless server with Nvidia and gnome-remote-desktop

Support
1

Hello,

I asked already in the gnome-remote-desktop repository and they told me this is
an issue of the distribution.

Systemdetails:

  • Hypervisor: Proxmox VE 9.3.2

  • GPU: One Nvidia Tesla T4 GPU, that my colleagues had in the storeroom from an earlier project, with Nvidia vGPU and Nvidia NLS DLS server

  • OS: ALMA 10.2

  • Gnome: 49.3 from the official OS repositories

  • Nvidia driver: 580.105.08

Expected behavior:

  • g-r-d and Gnome-Shell run on the Tesla T4

  • Overall performance increase, because the GUI stack is rendered on the GPU instead on the CPU

Current behavior:

  • g-r-d runs on the Tesla T4

  • Gnome-Shell does not and falls back to LLVMPIPE

Now to the details how I did set up this system:

  1. Followed the Proxmox VE Wiki to get the Nvidia GPU integrated into the hypervisor and assigned a vGPU profile to a VM

  2. Installed and configured the Nvidia NLS DLS appliance

  3. Installed ALMA 10.2 with profile “Server with GUI”

  4. Configured g-r-d to run as system wide process for headless login

  5. Installed the Nvidia GRID driver and configured it to get successfully a vGPU license

  6. Added in /etc/modprobe.d/blacklist.conf: blacklist nouveau

  7. Added in /etc/modprobe.d/blacklist.conf: options nouveau modeset=0

  8. Added in /etc/modprobe.d/blacklist.conf: options nvidia_drm modeset=1

    1. Without that option g-r-d and or Gnome would completely ignore the GPU with the message “Failed to open gpu ‘/dev/dri/card1’: DRM_CLIENT_CAP_UNIVERSAL_PLANES not supported”

  9. Following users got added to the groups “video” and “render”:

    1. myuser

    2. gnome-remote-desktop

    3. gdm

    4. pipwire

  10. Used “dracut --force” to add the modprobe configuration

  11. Edited the gnome-remote-desktop.service with “systemctl edit” and tried to disable CUDA:

    1. [Service]

    2. Environment=CUDA_VISIBLE_DEVICES=

  12. Rebooted the ALMA 10.2 VM

If I now connect through Windows MSTSC to the ALMA 10.2 VM the window for g-r-d credentials appear, which I fill out and get to the Gnome login page.

Here I enter the credentials of “myuser”, the system shows me after some seconds the default Gnome desktop and it is rendered through LLVMPIPE, because g-r-d is the only process nvidia-smi shows as running process.

Double bad is, that the LLVMPIPE stack crashes if I try use every Gnome application (Files, Terminal, etc.), only Firefox runs.

This is the full system log from the ALMA 10.2 VM since boot:

How do I reach the goal, to have g-r-d and Gnome-Shell fully rendered on the Tesla T4?

What am I missing?

Is this a driver issue and Nvidia wishes me just bad luck?

2

Fixed it myself.

Ignored one information my colleagues gave me and started to dig further:

  • Nvidia driver 595 is needed or else wayland-egl2 does not work and is needed
  • The package libva-nvidia-driver is needed, it installs libva for Nvidia video encode/decode
  • The GPU profiles added to a VM must be a 2Q-version (e.g. T4-2Q), because only the Q-Profiles have CUDA, CUDA is needed for HW acceleration through libva and 2GB of VRAM are needed or else the Nvidia driver fails to start CUDA
1 Like