Unable to Run Pylot & CARLA Simulator in WSL 2

[simonguozirui]

Here is my experience trying to set up Pylot using Windows Subsystem for Linux (WSL 2) and issues I met along the way. I hope this can be useful for people trying to do the same thing in the future.

TL;DR There seems to be issues with WSL or nvidia-docker that Pylot can't be run in WSL 2; would not recommend anyone trying to deploy Pylot on WSL 2 yet.

Motivation
I decided to try spinning up Pylot on WSL to utilize my GPU locally instead of running on servers, for easier development and potentially better performance. The recent NVIDIA's support for CUDA on WSL seems to be a possible way to get this working.

Spec and Set Up

• WSL Version: 2, runs Ubuntu 20.04.1 LTS.
• GPU: NVIDIA GeForce GTX 1060 with 6 GB VRAM
• Windows: Version 2004, OS build 21296.1010 (installed with Windows Insider Program, using Fast Ring/Dev Channel, do this at your own risk)
• X11 forwarding with VcXsrv in WSL 2 to show graphics; firewall config changed to allow forwarding.

I mostly followed the instructions using this CUDA on WSL User Guide on NVIDIA developer zone. Be very careful what drivers you are installing. This a pretty good guide and by following it I was able to:

• set up special driver and nvidia-docker in WSL using latest early preview Windows build
• able to run images (not Pylot) using nvidia-docker and get GPU acceleration working. Particular docker run --gpus all [nvcr.io/nvidia/k8s/cuda-sample:nbody](http://nvcr.io/nvidia/k8s/cuda-sample:nbody) nbody -gpu -benchmark
• set up X11 and VcXsrv to forward any graphics running in WSL to show up
• boot up Pylot docker image in WSL 2

However, the main issue is that the CARLA simulator cannot start successfully.

When I run nvidia-docker exec -i -t pylot /home/erdos/workspace/pylot/scripts/run_simulator.sh, the following output occurs and the program quits immediately.

4.24.3-0+++UE4+Release-4.24 518 0
Disabling core dumps.
sh: 1: xdg-user-dir: not found

Turns out these are normal warning flags from CARLA, but the simulator is unable to start successfully.

Since ./run_simulator.sh is running in headless state, no graphical window would open but it shouldn't be the case that it quits immediately.

Although the simulator cannot be run, the pylot.py scripts still can be run; it will just quit after timing out not finding a working simulator to connect to.

I tried a few things but none ended up working, including

• Running ./run_simulator.sh inside the container instead of running through nvidia-docker exec
• Running docker run --gpus and docker run -d -it --runtime=nvidia instead of using nvidia-docker as examples shown in the NVIDIA developer guide.
• Not deploy Pylot by Docker, but by manual installation.

I found similar issues on the CARLA repo, notably issues #2970 and issues #3431; it seems people (both using docker and without using docker) have no success with starting CARLA simulator in WSL yet and faces similar issues as I described. In fact, one of the CARLA's lead developers said they have not tested a WSL setup (here).

I believe this is an issue either with CARLA simulator, or nvidia-docker and WSL 2's support for CUDA in WSL. It is hard to debug as nvidia-smi does not work in WSL 2 yet (link). I can't tell whether the CARLA simulator running in docker was able to find the right GPU driver, and based on the Task Manager monitoring, the GPU was never active when I attempt to start the simulator.

Thanks @pschafhalter for help me debug this! Right now, I don't recommend anyone try run Pylot and CARLA simulator on WSL, as a lot of the WSL-GPU infrastructure has just been released and lack some essential tooling. If you found a way to resolve this, please share it with us!

[ICGog]

Closing this issue as we won't be able to fix it in the near future.

[NaveedBhuiyan]

I am trying to run the carla simulator on wsl2 also in a docker container but no luck. I keep getting the error "no vompatible vulkan found". However when I run the docker container on a linux host everything works just fine. Nvidia-smi works on wsl2 now but looks like it still cant find a compatible vulkan support specially the later carla releases. I am yet not sure which segment is causing the issue, wsl2 or carla itself

[simonguozirui]

@NaveedBhuiyan I tried to run Carla on WSL 2 (with nvidia-smi working) in docker too 2 years ago too and had very little luck. I would suggest just running it or the docker on Linux.