Skip to content

MPI Hello World with QNIBTerminal

Since I was asked (thanks Dmitry) via mail how to setup QNIBTerminal to run MPI jobs, I created a REAMDE within the qnib/compute repository, but why not put it in a blog post ( is Markdown, my blog is Markdown...)?

MPI HelloWorld

To fire up a small mpi script one needs

  • consul to bound tem all together (DNS wise and in regards of the slurm.conf)
  • slurmctld as master
  • at least two slurmd daemons to have a cluster (otherwise it would be only one :)


Fig was a project back when - the predecessor of docker-compose.

The fig file within the repository holds exactly this.

    image: qnib/consul

    image: qnib/slurmctld

    # inherits from qnib/slurmd
    image: qnib/compute

Let's fire it up and scale the nodes to 2.

$ fig up -d
Creating dockercompute_consul_1...
Creating dockercompute_slurmctld_1...
Creating dockercompute_compute_1...
$ fig scale compute=2
Starting dockercompute_compute_2...

Now we connect to the first node.

$ docker exec -ti dockercompute_compute_1 bash
[root@1df2666e8a45 /]#

After some time sinfo should show all nodes. The names are random, because I use the scale feature of fig (docker-compose), which is not able to use different hostnames (as to my knowledge).

$ sinfo
qnib*        up   infinite      2   idle 1df2666e8a45,f8d22e88943a

MPI the Thing

Now we compile the little mpi program hello_mpi.c.

$ mpicc -o /chome/cluser/hello_mpi /opt/qnib/src/hello_mpi.c
$ mpirun -n 1 /chome/cluser/hello_mpi
Process 0 on 1df2666e8a45 out of 1

To run it in parallel SLURM comes to the rescue and provides all the environment needed, we just have to run it.

$ salloc -N2 bash
salloc: Granted job allocation 2
$ mpirun /chome/cluser/hello_mpi
Process 7 on cdf11fef8b1e out of 8
Process 4 on cdf11fef8b1e out of 8
Process 6 on cdf11fef8b1e out of 8
Process 5 on cdf11fef8b1e out of 8
Process 3 on 647c9ef10a16 out of 8
Process 2 on 647c9ef10a16 out of 8
Process 0 on 647c9ef10a16 out of 8
Process 1 on 647c9ef10a16 out of 8