I wonder if there is a Dockerfile, or Singularity build for GeNN?

[russelljjarvis]

I wonder if there is a Dockerfile, or Singularity build for GeNN+CUDA tools? I am wondering this as I am have been debugging my GeNN installation regular GeNN works with CUDA but PyGeNN cannot build a CUDA backend, and so it doesn't have one.

Even if I succeed at debugging my laptop installation of PyGeNN I will ultimately have to redo the installation on HPC, and without administrative privileges.

I have noticed that a lot of HPC projects use the singularity to do simplify and containerize complex builds on HPC. Does something like this already exist for GeNN+PyGeNN?

[neworderofjamie]

What OS is your laptop running and how does PyGeNN fail? If C++ GeNN works ok, it shouldn't be too tricky to install - definitely no administrative access is required and, between us, we've installed it on several HPC systems without too much pain.

Creating a Dockerfile for GeNN is on my todo list but you would be more than welcome to help me out with it! The idea would be to have a docker image with an entrypoint that supports the kind of functionality of NEST's to launch interactive sessions and notebook environments (https://github.com/nest/nest-docker/blob/master/src/3.3/entrypoint.sh#L22-L50).

[russelljjarvis]

I kind of think for Windows users Docker might be better than WSL2. Another issue with WSL2 is it doesn't have access to all the system RAM by default.

[neworderofjamie]

I have successfully got WSL working on Windows 11 with CUDA interop but not so far on Windows 10 (https://docs.nvidia.com/cuda/wsl-user-guide/index.html were what I followed). Can you now build Windows docker images? If so, that might well be better and would provide a more familiar shell for Windows users

[russelljjarvis]

Creating a Dockerfile for GeNN is on my todo list but you would be more than welcome to help me out with it! The idea would be to have a docker image with an entrypoint that supports the kind of functionality of NEST's to launch interactive sessions and notebook environments (https://github.com/nest/nest-docker/blob/master/src/3.3/entrypoint.sh#L22-L50).

Because of multistage builds, I think you could do:

FROM docker-registry.ebrains.eu/nest/nest-simulator:3.3

Somewhere near the end of the Dockerfile code I provided. That would cause some bloat, another approach would be to install

FROM jupyter:scipynotebook # not a precise image name, just a gist.

[russelljjarvis]

I have successfully got WSL working on Windows 11 with CUDA interop but not so far on Windows 10 (https://docs.nvidia.com/cuda/wsl-user-guide/index.html were what I followed). Can you now build Windows docker images? If so, that might well be better and would provide a more familiar shell for Windows users

Yes you can build windows Dockerimages, but its not something I have ever thought to do.
https://hub.docker.com/_/microsoft-windows-base-os-images

[russelljjarvis]

I have successfully got WSL working on Windows 11 with CUDA interop but not so far on Windows 10 (https://docs.nvidia.com/cuda/wsl-user-guide/index.html were what I followed). Can you now build Windows docker images? If so, that might well be better and would provide a more familiar shell for Windows users

-- Oh thanks for this you reminded me to check my host CUDA, I think I destroyed the host cudatools environment recently.

Yes you can build windows Dockerimages, but it's not something I have ever thought to do. https://hub.docker.com/_/microsoft-windows-base-os-images

For the idea of a based on the NEST container where there is an Jupyter notebook entry point, in that case, it might not matter that much what the Docker OS is.

I often like to do things like !pwd; !ls; or !wget inside notebooks, so I always naturally gravitate to Linux. Apparently powershell is POSIX compliant now, so occasionally GNU core util commands work in powershell.

[russelljjarvis]

I have made and tested out this Dockerfile

FROM nvidia/cuda:11.8.0-devel-ubuntu22.04
ENV CUDA_PATH=/usr/local/cuda-11.8 
# RUN cuda-memcheck ./has_cuda ; exit 0
# RUN nvcc; exit 0
RUN apt update; 
RUN apt install git build-essential lcov wget -y --no-install-recommends 
ENV CONDA_DIR /opt/conda
RUN wget --quiet https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda.sh && \
     /bin/bash ~/miniconda.sh -b -p /opt/conda

ENV PATH=$CONDA_DIR/bin:$PATH
WORKDIR $HOME/workdir
RUN $(which conda) install -c miniconda swig numpy -y
RUN git clone https://github.com/genn-team/genn.git
RUN mv genn GeNN
WORKDIR GeNN
ENV PATH=$PATH:$HOME/workdir/GeNN/bin
RUN make DYNAMIC=1 LIBRARY_DIRECTORY=`pwd`/pygenn/genn_wrapper/
RUN python setup.py develop
WORKDIR $HOME/workdir/GeNN/userproject/PotjansMicrocircuit_project
RUN make
RUN ./generate_run test 
WORKDIR $HOME/workdir/GeNN/
RUN bash tests/run_tests.sh; exit 0 # This would fail any way for reasons outside of my control.
ENTRYPOINT ["/bin/bash"]

It fails at the step
RUN ./generate_run test
but I think it is close.

Maybe I need a different Docker command for building with access to system cuda?

[neworderofjamie]

Thank you for this - I am actually going to a hackathon this afternoon to try and get some GeNN collaborators so I might try and get some more hands on this

[russelljjarvis]

I wonder if you have participated in BrainHack before? Its a good hackathon with international and local components. It used to be more focused on fMRI/EEG, but electrical neuronal network modelling is creeping in.

You can contribute a project by making an issue for the repo at your local chapter.
https://github.com/brainhackorg/global2022
This is my one:
brainhackorg/global2022#17

[neworderofjamie]

The docker file in Genn's docker2 branch works for me. With the nvidia docker stuff installed, you just need to run the container with docker run --gpus all to access the GPUs - will tidy it up tomorrow

[russelljjarvis]

Sweet. It's minimal. Great I love it! https://github.com/genn-team/genn/blob/docker2/Dockerfile
I will run it in a bit, thanks!

[russelljjarvis]

Maybe it was actually conda that was stuffing everything up in my build.

[neworderofjamie]

As of the 4.8.0 release, there is now an official GeNN docker image built from https://github.com/genn-team/genn/blob/master/Dockerfile and with prebuilt versions published at https://hub.docker.com/repository/docker/gennteam/genn