Better Performance - Force rviz to run on second GPU or utilize more resources?

[rwbot]

I'm viewing multiple layers of dense pointclouds in rviz, and it's lagging with just one layer, and gets worse as more layers are activated. I have two RTX 3080s, and at no point will GPU 1 get utilized, only GPU 0, usually around 50%, while CPU load is rarely more than 20% and RAM goes to 25GB max.

Is there any way to improve performance by forcing rviz to only use a specific graphics card or utilize more resources?

Your environment

• OS Version: e.g. Ubuntu 20.04
• CPU: AMD® Ryzen threadripper pro 3995wx 64-cores × 128
• GPU: 2 x NVIDIA GeForce RTX 3080/PCIe/SSE2
• RAM: 128GB
• ROS Distro: [Noetic]
• RViz, Qt, OGRE, OpenGl version as printed by rviz:

[${node}.VisualizerApp::init]: rviz version 1.14.10
[${node}.VisualizerApp::init]: compiled against Qt version 5.12.8
[${node}.VisualizerApp::init]: compiled against OGRE version 1.9.0 (Ghadamon)
[/rviz_1637200062035959003.RenderSystem::forceGlVersion]: Forcing OpenGl version 0.
[/rviz_1637200062035959003.RenderWindow* rviz::RenderSystem::makeRenderWindow]: Stereo is NOT SUPPORTED
[/rviz_1637200062035959003.RenderSystem::detectGlVersion]: OpenGL device: NVIDIA GeForce RTX 3080/PCIe/SSE2
[/rviz_1637200062035959003.RenderSystem::detectGlVersion]: OpenGl version: 4.6 (GLSL 4.6).

[rhaschke]

Basically, I don't think there is a way to utilize both GPUs. Eventually, you have a single rviz window only.
When you report 20% CPU usage, is that averaged over all 128 threads? I guess a few of them are at 100%?
Essentially, PointCloud messages are processed by a single (non-GUI) thread. Additionally, there is the GUI/main thread doing all rendering and GUI interaction. To improve your situation, you first need to understand where the bottleneck actually is.
What do you exactly mean by PointCloud layers? Do you have various PC displays listening to different topics?

I recently changed the message processing for PointCloud displays (#1635) to use a single message queue and processing thread for all PC displays in order to fix a memory leak. You could try to revert that commit and see whether it improves your situation.