using MPI with docker containers

Question

I have created a docker image based on Ubuntu 16.04 and with all the dependencies needed to run MPI.

It is public on docker-hub at: https://hub.docker.com/r/orwel84/ubuntu-16-mpi/

I use this image to create an MPI container. I can also compile a simple mpi-hello-world.c (which comes inside the container) and run it with mpirun.

These are the steps I use and, (if you have Docker installed you can reproduce them too) :

1. docker run -it orwel84/ubuntu-16-mpi bash
2. (on container's shell) mpirun -np 4 --allow-run-as-root ./mpi_hello_world

You will see Output:

Hello world from processor 6f9b11cef939, rank 0 out of 4 processors

Hello world from processor 6f9b11cef939, rank 1 out of 4 processors

Hello world from processor 6f9b11cef939, rank 2 out of 4 processors

Hello world from processor 6f9b11cef939, rank 3 out of 4 processors

Question:

Right now all the above four mpi processes run inside a single container.

How can I use mpirun to run on multiple containers on a single host? And also how can I use Docker swarm to run on multiple nodes of the swarm?

Please help. Thankyou.

Entoil

How could u solve the problem finally? – Starnes

Generally, it's a good idea to run a single container per node (aka Fully contained mode) - Here's a link ibm.com/docs/en/smpi/… – Linkboy

Answer 1

for multiple containers in a single machine:

1.You can create multiple containers on a single host machine and inspect their ip address (under docker bridge network).

docker inspect -f "{{ .NetworkSettings.IPAddress }}" containerName

2. Now attach to one of the containers and create a host file listing ip-addresses of the containers you have created.

3.Now run the application:

mpirun -np 4 -hostfile hostfile_you_created ./mpi_hello_world

Circlet

Is there a reason this would be faster than running all the processes in one docker container? – Speciality

I don't think the speed will be faster in this way. The docker overhead is negligible most of the cases. Multiple containers are useful when one machine can not host all the ranks and that needs to be split across multiple containers across hosts. But in that situation containers need to share the same network (like swarm provides overlay), so they can reach each other. – Circlet