CARLA Multi-GPU Camera-Based Sensors Frame Drop

[NesquikPlus]

Setup

• CARLA version: 0.9.15.2
• Platform: Ubuntu 20.04.6 LTS
• Python version: 3.8
• GPU: 2 RTX 4060's
• GPU Drivers: 580.95.05

Describe the bug

When running CARLA in multi-GPU mode, sensors intermittently fail to produce frames during synchronous simulation if the followings are met:
--> There is at least one GPU-based sensor (rgb camera, depth camera or any other custom sensor that uses GPU.)
--> At least one of the sensors has the attribute 'sensor_tick' set. This is not necessarily GPU-based sensor.

If the above conditions are met, approximately every 50–100 simulation ticks, all sensors simultaneously drop a frame — i.e., no listen() callback is triggered, and no new data is emitted for that world tick.

Steps to reproduce
You may try to run the python script I provided or write a synchronous client script and add two RGB cameras and one GNSS sensor with at least one of them having the attribute 'sensor_tick' set.

Expected behavior
I'd expect similar behavior to running CARLA in single GPU mode. No frame drops.
Similarly when none of the sensors have 'sensor_tick' attribute set there are no frame drops.

Logs
Client script logs:
output.txt

Scripts
Client script:
github_issue.py

[PatrickFSLin]

Following this topic, I’d also like to ask the following question:
Can CARLA’s Multi-GPU mode run on a single machine that has two or more graphics cards?

Thanks

[NesquikPlus]

Yes it can. @PatrickFSLin.
That's why we specify r.GraphicsAdapter starting the secondary servers:
https://carla.readthedocs.io/en/latest/adv_multigpu/