Skip to content

NAG-DevOps/docker-gpaw-openmpi

 
 

Repository files navigation

main

Description

Dockerfile for GPAW https://wiki.fysik.dtu.dk/gpaw/ built against openmpi, based on Fedora.

The docker images are available at dockerhub

Usage

First, make sure you are able to run the docker run hello-world example https://docs.docker.com/get-started/. Note on MS Windows install https://hub.docker.com/editions/community/docker-ce-desktop-windows.

Then test the basic GPAW functionality

docker run --rm -it marcindulak/gpaw-openmpi:latest bash -c '. /etc/profile.d/modules.sh&& module use /usr/share/modulefiles&& module load mpi/openmpi && PYTHONPATH=$MPI_PYTHON3_SITEARCH OMP_NUM_THREADS=1 mpiexec --allow-run-as-root -np 2 python3 -c "import gpaw.mpi; print(gpaw.mpi.rank)"'

Note: if on MS Windows you are getting 'image operating system "linux" cannot be used on this platform' follow https://docs.docker.com/docker-for-windows/#switch-between-windows-and-linux-containers

When running GPAW jobs inside of the container you want the output files created by GPAW to be accessible locally on the machine running the job. In order to achieve this create a local storage directory for this particular myjob.

mkdir myjob

Create a GPAW input script and save into the myjob directory.

Then start the job mounting the local storage directory as a docker volume https://docs.docker.com/storage/volumes/ inside of the container. You have two choices, listed below.

Run a job with docker compose

docker compose -f docker-compose.myjob.yml up --exit-code-from gpaw

Remove the terminated container

docker compose -f docker-compose.myjob.yml down

Run a job manually

docker run --name myjob --rm -it -v "$(pwd)/myjob:/mnt" marcindulak/gpaw-openmpi:latest bash -c '. /etc/profile.d/modules.sh&& module use /usr/share/modulefiles&& module load mpi/openmpi && cd /mnt&& PYTHONPATH=$MPI_PYTHON3_SITEARCH OMP_NUM_THREADS=1 mpiexec --allow-run-as-root -np 2 python3 h2.py'

Examine the created output file

cat myjob/h2.txt

Building

Locally

Build based on the local Dockerfile

docker build -t gpaw-openmpi .

List images

docker images

Releasing

Docker image tagging convention

The docker image is tagged with the GPAW upstream version appended by a build number, for example the first docker image of GPAW version 1.4.0 is tagged with 1.4.0-1. Note that this docker image may correspond to e.g. Fedora's RPM 1.4.0-10.fc30. See the ENV values in Dockerfile.

Build and test the image, see buildx.

Create a builder instance

docker buildx create --name node-docker-gpaw-openmpi --platform linux/amd64,linux/arm64
docker buildx use node-docker-gpaw-openmpi

Build the image for supported targets

BUILDKIT_PROGRESS=plain docker buildx bake -f docker-compose.test.yml --set '*.platform=linux/amd64,linux/arm64' --no-cache
# or
# BUILDKIT_PROGRESS=plain docker buildx build --platform linux/amd64,linux/arm64 --no-cache .

Load the machine target image for testing. Only one platform can be loaded, see docker/buildx#59 and https://stackoverflow.com/questions/71765232/what-is-default-platform-type-in-docker-image

docker buildx build --platform $(docker system info --format '{{.OSType}}/{{.Architecture}}') -t docker-gpaw-openmpi-sut:latest --load .

Test the image

docker compose -f docker-compose.test.yml up --exit-code-from sut

Tag and push the images (need to build again, using --push option docker/buildx#1152)

BUILDKIT_PROGRESS=plain docker buildx build --platform linux/amd64,linux/arm64 -t marcindulak/gpaw-openmpi:1.4.0-1 --push .
BUILDKIT_PROGRESS=plain docker buildx build --platform linux/amd64,linux/arm64 -t marcindulak/gpaw-openmpi:latest --push .

Remove the builder instance

docker buildx stop node-docker-gpaw-openmpi
docker buildx rm node-docker-gpaw-openmpi

Dependencies

docker and optionally docker compose

License of this Dockerfile

Apache-2.0

Note please consult the GPAW software https://gitlab.com/gpaw/gpaw/blob/master/LICENSE

Todo

About

GPAW Openmpi Dockerfile based on Fedora

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Dockerfile 79.9%
  • Python 20.1%