Use GPU Acceleration for Graphical Applications

This guide explains how to enable GPU hardware acceleration for applications in a remote desktop (VNC) or local graphical environment to improve rendering performance. It is suitable for scenarios such as 3D design, scientific visualization, video processing, and CAD/CAM that require high-performance graphics processing.

Technical Architecture

VirtualGL provides GPU acceleration through two technical approaches:

Rendering Path	Advantages	Disadvantages	Recommended Scenarios
GLX backend acceleration	Strong compatibility, supports most 3D applications	More complex configuration and higher system overhead	Traditional 3D applications and environments with high compatibility requirements
EGL backend acceleration	Simpler configuration, lower system overhead, better performance	Limited compatibility with some legacy 3D applications	Modern applications, containerized environments, and cloud platforms

Prerequisites

• The required GPU driver has been installed.

Method 1: GLX Backend Accelerated Rendering

Configure the X Server

NVIDIA GPU configuration steps:

1. Generate the Xorg configuration:

   sudo nvidia-xconfig --enable-all-gpus --separate-x-screens

2. Query the GPU BusID:

   lspci | grep NVIDIA
   # Example output: 00:07.0 3D controller: NVIDIA Corporation GP104GL [Tesla P4]

3. Edit the Xorg configuration:

   sudo vi /etc/X11/xorg.conf

   Add BusID to the corresponding Device section:

   Section "Device"
       Identifier     "Device0"
       Driver         "nvidia"
       BusID          "PCI:0:7:0"  # Adjust according to the actual BusID
   EndSection

4. Reboot the system:

   sudo reboot

Install a Display Manager

Multiple display managers are supported, such as GDM, KDM, and LightDM. The following example installs LightDM on Ubuntu 18.04:

Ubuntu/Debian systems (LightDM):

# Install LightDM
sudo apt update
sudo apt install lightdm

# Set it as the default display manager
sudo dpkg-reconfigure lightdm

# Or enable it by command
sudo systemctl enable lightdm

# Restart the display service
sudo systemctl restart lightdm

Configure the VirtualGL Server

Run the configuration script:

sudo /opt/VirtualGL/bin/vglserver_config

Configuration options:

Please choose a configuration option:
1) Configure server for use with VirtualGL (GLX + EGL back ends)
2) Configure server for use with VirtualGL (EGL back end only)
X) Exit

Choose 1 (GLX + EGL)

Restrict 3D X server access to the vglusers group?
[Y/n] n
# Choose n: all users can access it; choose Y: only members of the vglusers group can access it

Restrict framebuffer device access to the vglusers group?
[Y/n] n

Disable the XTEST extension?
[Y/n] n  # You must choose n, otherwise remote desktop operations are affected

Restart the display service:

sudo systemctl restart lightdm  # or gdm
# or reboot the system
sudo reboot

Verify the Configuration

Check the X Server configuration:

# Check the display configuration
xdpyinfo -display :0

# Check the OpenGL configuration
/opt/VirtualGL/bin/glxinfo -display :0 -c

Expected results:

• The list of available visual effects should be displayed.
• No error messages should be shown.
• The GPU should be detected correctly.

Run a 3D Application

Run after connecting through VNC:

# Basic usage
vglrun ./my_3d_app

# Explicitly specify the display device (optional)
vglrun -d :0 ./my_3d_app

VNC session example:

# Run in a new terminal after connecting to VNC
vglrun glxgears  # Test application

Method 2: EGL Backend Accelerated Rendering

Advantages and Limitations of EGL

Advantages:

• Lower system overhead
• Better performance
• Simpler configuration

Limitations:

• The application must support EGL. Some legacy applications may not be compatible.

Configure the EGL Backend

Run the configuration script:

sudo /opt/VirtualGL/bin/vglserver_config

Configuration options:

Choose option 2 (Configure server for use with VirtualGL (EGL back end only))

Restrict framebuffer device access to the vglusers group?
[Y/n] n  # Choose n to allow all users to access it

Verify the EGL Configuration

# Check EGL support
/opt/VirtualGL/bin/eglinfo egl0

# Expected result: the EGL configuration list is displayed with no errors

Run a 3D Application in EGL Mode

Using environment variables:

# Method 1: use an environment variable
export VGL_DISPLAY=egl
vglrun ./my_3d_app

# Method 2: set it inline for a single command
VGL_DISPLAY=egl vglrun ./my_3d_app

Using command-line parameters:

# Use the -d parameter to specify the EGL device
vglrun -d egl0 ./my_3d_app

Multi-GPU Environment Configuration (To Be Verified)

GPU Detection and Selection

View available GPUs:

# NVIDIA
nvidia-smi -L

# AMD
rocm-smi --showproductname

# Generic method
lspci | grep -E "(VGA|3D)"

Multi-GPU Configuration for the GLX Backend

Different X Server configurations:

# Start an additional X Server. Make sure the port does not conflict with the VNC session.
Xorg :10 -ac -br -noreset -sharevts -logfile /var/log/Xorg.10.log -config /etc/X11/xorg.conf.gpu1

# Use a specific GPU
vglrun -d :10 ./app

Multi-GPU Configuration for the EGL Backend

Using device paths:

# Select by DRI device path
VGL_DISPLAY=egl vglrun -d /dev/dri/card0 ./app
VGL_DISPLAY=egl vglrun -d /dev/dri/card1 ./app

Using EGL device IDs:

# Select by EGL device ID
vglrun -d egl0 ./app  # GPU0
vglrun -d egl1 ./app  # GPU1
vglrun -d egl2 ./app  # GPU2

Frequently Asked Questions

Q: Can one GPU card be used by multiple users or processes at the same time?

A: Yes. This can be achieved through software-level resource sharing or hardware-level partitioning and isolation through GPU virtualization such as MIG. Choose the appropriate approach based on job requirements and isolation needs.

Q: How can I prevent multiple users or processes from concentrating work on only a few GPUs, such as gpu0-10?

A: In FCP scenarios, GPU resource isolation is generally implemented by binding users to login nodes. In cluster planning, different login nodes and their GPUs are already assigned to specific users or groups. This mapping is configured and maintained centrally by the system administrator.