testmpi

HOME About us License Miscellaneous

PCIC

Parallel Computing in Container

Testing MPI implementations inside Docker / Singularity container

This package tests various implementations of MPI inside Singularity containers derived directly from Docker images. Two major implementations are tested here: MPICH and OpenMPI. Package contents:

Source code mpi_check.c and Makefile
Python script master.py to automatically create Docker images of various MPI implementations, as well as samples of job scheduling batch files for an HPC system (*.sbatch).

MPICH

It has ABI (Application Binary Interface) compatibility across multiple versions and vendors. More information can be found here: https://www.mpich.org/abi/

OpenMPI

Compatibility is guaranteed within a major version number, where a version number is given as: major.minor.release. There are a number of versions:

Current at the time of writing (December 2019): 4.x
Still supported by OpenMPI: 3.x
No longer supported: 2.x and 1.x More information on this can be found here: https://www.open-mpi.org/software/ompi/versions/

How to use

Using a pre-built Docker image

This is the simplest way.

Pull one of these pre-built Docker images from DockerHub and turn it into Singularity container:

singularity pull docker://lahur/testmpich
singularity pull docker://lahur/testopenmpi-4.0.2
singularity pull docker://lahur/testopenmpi-3.1.4
singularity pull docker://lahur/testopenmpi-2.1.6

Run the command within the container For example, on Pearcey, depending on the MPI implmentation, run one of these batch files.

sbatch pearcey-mpich.sbatch
sbatch pearcey-openmpi-4.0.2.sbatch
sbatch pearcey-openmpi-3.1.4.sbatch
sbatch pearcey-openmpi-2.1.6.sbatch

Building Docker images automatically

The Python script master.py can build a number of Docker images automatically. Inside the file, at the top, there’s a section for user settings:

mpi_targets = ["mpich", "openmpi-4.0.2", "openmpi-3.1.4", "openmpi-2.1.6"]

target_prepend = "lahur/test"
target_append = ":latest"

dry_run = True

HPC = "pearcey"

Note

mpi_targets is a list of MPI targets. Modify this to build different target.
The name of Docker image created will be in this format: target_prepend target target_append. For example: lahur/testopenmpi-4.0.2:latest.
Set dry_run to False if you want to actually build the image. Otherwise only Dockerfiles and sample batch files are produced.
HPC is the name of HPC that will be prepended to sample batch files.

Executing the script is simple.

python master.py

The script will create the necessary Dockerfiles, build the images, and make samples of SLURM batch files.

Building Docker image manually

To build Docker image manually, you can modify the existing Dockerfiles. Then run the build command.

docker build -t my_dockerHub_account/my_docker_image:image_tag -f my_dockerfile .

If you wish to use the image on another machine, push the Docker image to your account in DockerHub.

docker push my_dockerHub_account/my_docker_image:image_tag

Then pull it from the target machine as a singularity object.

singularity pull docker://my_dockerHub_account/my_docker_image:image_tag