Distributing jobs across nodes

In order to distribute MPI jobs across the nodes in a cluster, one has to either specify these nodes as a parameter to the mpirun/mpiexec command or make use of a host file. This host file contains the names of the nodes on the network which will be available for a run, along with the number of available slots on the host.

A prerequisite for running MPI applications on a remote node is that the MPI runtime is installed on that node, and that password-less access has been configured for that node. This means that so long as the master node has the SSH keys installed, it can log into each of these nodes in order to launch the MPI application on it.

Setting up an MPI node

After installing MPI on a node, the next step is to set up password-less SSH access for the master node. This requires the SSH server to be installed on the node (part of the ssh package on Debian-based distributions). After this we need to generate and install the SSH key.

One way to easily do this...