Now that you have a small cluster let’s test that it works.
MPI relies on the same path for files on all nodes, the easiest way to do this is a shared folder between all the nodes, which for this tutorial was configured as /mnt/mpi_shared.
bryandixon@ccsc1:~/CCSC-Tutorial$ cd /mnt/mpi_shared
bryandixon@ccsc1:/mnt/mpi_shared$
Now let’s grab a hello world mpi program from an MPI tutorial
bryandixon@ccsc1:/mnt/mpi_shared$ wget https://raw.githubusercontent.com/mpitutorial/mpitutorial/gh-pages/tutorials/mpi-hello-world/code/mpi_hello_world.c
Now we have the MPI code.
Now let’s compile our MPI code:
bryandixon@ccsc1:/mnt/mpi_shared$ mpicc -o mpi_hello_world mpi_hello_world.c
bryandixon@ccsc1:/mnt/mpi_shared$ ls
mpi_hello_world mpi_hello_world.c
Now we have a mpi_hello_world executable. To be able to run the MPI program we need to be able to SSH to all our nodes without password or other prompts. We have configured hostnames for our 3 nodes of gcp1, gcp2, and gcp3 so ssh into all 3 of these nodes to make sure there are no prompts any more first.
Now create a hosts file for MPI:
bryandixon@ccsc1:/mnt/mpi_shared$ printf "gcp1\ngcp2\ngcp3\n" > hosts
bryandixon@ccsc1:/mnt/mpi_shared$ cat hosts
gcp1
gcp2
gcp3
This file tells mpi which nodes to run on. Now we can run the code:
bryandixon@ccsc1:/mnt/mpi_shared$ mpirun -n 3 --hostfile hosts ./mpi_hello_world
Hello world from processor ccsc1, rank 0 out of 3 processors
Hello world from processor ccsc2, rank 1 out of 3 processors
Hello world from processor ccsc3, rank 2 out of 3 processors
We now successfully have a cluster and are able to run MPI code on it.