CP2K docker containers are suited for many host systems. A convenient way to use them is to run such a container with Apptainer which is described in the following sections. Further down, you will find the information how to run and build CP2K production containers with docker.
It is possible to download and run CP2K docker containers with Apptainer as .sif files. Such Apptainer files are especially suited to be run on HPC systems or systems where docker is not installed.
Install the latest Apptainer version as described here. Most Linux distributions provide rpm packages for Apptainer.
The available pre-built CP2K production docker containers can be found here. The name of a docker container indicates CP2K version, MPI implementation (MPICH or OpenMPI), target CPU (generic, haswell, or skylake-avx512), CUDA support, and cp2k binary version (e.g. psmp for MPI parallelized with OpenMP support). If you don't know your CPU model, then try the CPU target generic (NEHALEM) first or simply choose latest:
apptainer pull docker://cp2k/cp2k:latest
which will download the mpich_generic_psmp image of the latest CP2K release version.
Check if the pre-built container runs on your system with
apptainer run cp2k_latest.sif cp2k -h -v
or simply
cp2k_latest.sif
should work as well, if the .sif file is executable. This test should not give any error messages. A warning like setlocale: LC_ALL: cannot change locale will disappear when LC_ALL=C is set in advance:
LC_ALL=C cp2k_latest.sif cp2k
If the .sif is not executable, then you can use the command chmod a+x to make it executable:
chmod a+x cp2k_latest.sif
A more extensive testing, a kind of self-test, can be performed with the run_tests command
apptainer run -B $PWD:/mnt cp2k_latest.sif run_tests
which launches a full CP2K regression test run within the container. The test run will use 8 tasks (CPU cores) by default. If you have more or less CPU cores available, you can set the maximum number of tasks (CPU cores), preferibly a multiple of 4 for this test, with the --maxtasks flag
apptainer run -B $PWD:/mnt cp2k_latest.sif run_tests --maxtasks 32
CUDA enabled containers can be employed on host systems which provide also NVIDIA GPU resources. Pull a CUDA enabled CP2K docker container matching the GPU type of your system, e.g. for Pascal GPUs (P100)
apptainer pull docker://cp2k/cp2k:2024.1_mpich_generic_cuda_P100_psmp
Run the container with the --nv flag using the nvidia-smi command from the container
apptainer run --nv ./cp2k_2024.1_mpich_generic_cuda_P100_psmp.sif nvidia-smi
and compare its output with the ouput of the nvidia-smi command from the host system. These outputs should agree when the container is able to recognize the GPU resources of the host system, correctly. Check also the CUDA version of the host system shown (top right) in the output of nvidia-smi command. That CUDA version should match or be newer than the CUDA version installed in the container. Run the command
apptainer run --nv ./cp2k_2024.1_mpich_generic_cuda_P100_psmp.sif nvcc --version
to retrieve the CUDA version of the actual container. With the run_tests command
apptainer run --nv -B $PWD:/mnt ./cp2k_2024.1_mpich_generic_cuda_P100_psmp.sif run_tests
one can check eventually if the container works correctly and the GPU usage can be monitored on the host system with the nvidia-smi command while running.
The MPI of the container can be employed to run CP2K within a compute node, e.g. using 4 MPI ranks with 2 OpenMP threads each
apptainer run -B $PWD cp2k_2024.1_mpich_generic_psmp.sif mpiexec -n 4 -genv OMP_NUM_THREADS=2 cp2k -i H2O-32.inp
with MPICH and similarly with OpenMPI
apptainer run -B $PWD cp2k_2024.1_openmpi_generic_psmp.sif mpiexec -n 4 -x OMP_NUM_THREADS=2 cp2k -i H2O-32.inp
For multi-node runs on HPC cluster systems, it is required to use the MPI of the host system
export OMP_NUM_THREADS=2
mpiexec -n 4 apptainer run -B $PWD cp2k_2024.1_mpich_generic_psmp.sif cp2k -i H2O-32.inp
to achieve best performance, but incompabilities, e.g. because of proprietary drivers or installations, might disable runing the pre-built container with the host MPI. If the host system has installed SLURM as a scheduler, srun can (should) be used instead of mpiexec (or mpirun)
srun -n 4 apptainer run -B $PWD cp2k_2024.1_mpich_generic_psmp.sif cp2k -i H2O-32.inp
With SLURM, srun is usually the proper way to launch a production run in batch mode using a CP2K sif file.
The containers provide also various CP2K tools like dumpdcd, graph, and xyz2dcd. Run
cp2k_latest.sif dumpdcd -h
for more information and see below.
Install the latest docker version on the host system. Most Linux distributions provide rpm packages for docker.
Pre-built CP2K docker containers can also be downloaded from Dockerhub using docker. There are containers compiled using MPICH and OpenMPI for various target CPUs like generic (NEHALEM) and for the x86_64 Intel CPUs haswell and skylake-avx512 are available.
Check if the pre-built container runs on your system using docker, e.g. with
docker run -it --rm cp2k/cp2k:latest
or
docker run -it --rm cp2k/cp2k:latest cp2k -h -v
This test should not give any error messages. A more extensive testing can be performed with run_tests
docker run -it --rm --shm-size=1g -v $PWD:/mnt -u $(id -u $USER):$(id -g $USER) cp2k/cp2k:latest run_tests
The flag -it keeps the interactive terminal attached which allows for an interrupt of the run with Ctrl-C. The --rm flag removes automatically the created container when it exits. Furthermore, the --shm-size=1g flag is needed to increase the shared memory size for the MPI inter-process communication.
The MPI of the container can be employed to run CP2K within a compute node. The containers built with MPICH can be run with
docker run -it --rm --shm-size=1g -v $PWD:/mnt -u $(id -u $USER):$(id -g $USER) cp2k/cp2k:2024.1_mpich_generic_psmp mpirun -n 4 -genv OMP_NUM_THREADS=2 cp2k -i H2O-32.inp
whereas the containers built with OpenMPI can be run, e.g. using
docker run -it --rm --shm-size=1g -v $PWD:/mnt -u $(id -u $USER):$(id -g $USER) cp2k/cp2k:2024.1_openmpi_generic_psmp mpirun -bind-to none -np 4 -x OMP_NUM_THREADS=2 cp2k -i H2O-32.inp
After pulling a CUDA enabled CP2K docker container, e.g. for Pascal GPUs (P100), similarly to the section above, the following docker commands will check
docker run -it --rm --gpus all cp2k/cp2k:2024.1_mpich_generic_cuda_P100_psmp nvidia-smi
docker run -it --rm --gpus all cp2k/cp2k:2024.1_mpich_generic_cuda_P100_psmp nvcc --version
docker run -it --rm --gpus all --shm-size=1g -v $PWD:/mnt -u $(id -u $USER):$(id -g $USER) cp2k/cp2k:2024.1_mpich_generic_cuda_P100_psmp run_tests
if the container is working correctly.
All containers provide as further binaries:
cp2k.popt(enforcesOMP_NUM_THREADS=1by contrast tocp2k.psmpor justcp2k)cp2k_shell(runs the CP2K shell tool)dumpdcd(processing and conversion of.dcdfiles)xyz2dcd(convert.xyzfiles to.dcdfiles and optionally include cell information from CP2K's.cellfile)graph(FES post-processing tool for Metadynamics runs)
You can build your own CP2K docker container, e.g. using the Docker file 2024.1_mpich_generic_psmp.Dockerfile for the CP2K version 2024.1, by running
docker build -f ./2024.1_mpich_generic_psmp.Dockerfile -t $USER/cp2k:2024.1_mpich_generic_psmp .
or for the current CP2K trunk version (see master branch) with
docker build -f ./master_mpich_generic_psmp.Dockerfile -t cp2k/cp2k:master$(date +%Y%m%d)_mpich_generic_psmp .
Each Docker file includes a Usage line in its header with the docker build command to run that file.
If the docker files will include any CP2K regression test during its build, the log file size will be much larger. In that case, it is recommended either to use docker's legacy builder
DOCKER_BUILDKIT=0 docker build -f ./2024.1_mpich_generic_psmp.Dockerfile -t cp2k/cp2k:2024.1_mpich_generic_psmp .
or to add the --progress flag when using the new docker builder buildx, e.g.
docker buildx build --progress=plain -f ./2024.1_mpich_generic_psmp.Dockerfile -t cp2k/cp2k:2024.1_mpich_generic_psmp .
The Python script generate_docker_files generates all docker files in this folder if run without any further option. It provides a few command line options allowing for an adaptation of the created docker files to the actual needs, e.g. with
./generate_docker_files.py -j 32 --test
it will create docker files using 32 instead of the default 8 CPU cores for building the docker containers and the container will run a full CP2K regression test after the container has been built. Run generate_docker_files -h for more details about the available options.