CODEWITHSUNDEEP
openmpi
Andre
Andre

Reputation: 558

What command does MPI run over SSH?

I have OpenMPI installed and I'm running a script on multiple nodes with mpiexec. OpenMPI requires that the node I'm running the mpiexec command on have SSH access to the other nodes in the cluster.

What specifically does OpenMPI do over SSH to start the processes on the other nodes? It runs my MPI script, sure, but how does MPI run in such that each node is assigned a rank, for example?

Thank you.

Upvotes: 1

Views: 4917

Answers (1)

Gilles Gouaillardet
Gilles Gouaillardet

Reputation: 8395

Unless you are running under a (supported) resource manager (such as Slurm, PBS or other), the plm/rsh component will be used to start the MPI app.

Long story short, Open MPI uses a distributed virtual machine (DVM) to launch the MPI tasks. The first step is to have one daemon per node. The initial "daemon" is mpirun, and then one orted daemon have to be remotely spawned on each other node, and this is where plm/rsh uses SSH.

By default, if you are running on less than 64 nodes, then mpirun will SSH to all the other nodes. But if you are running on a larger number of nodes, then mpirun will use a tree spawn algorithm, in which other nodes might ssh to other nodes. Bottom line, if you are using ssh with Open MPI, and unless you are running on a small cluster with default settings, all nodes should be able to ssh passwordless to all nodes.

Upvotes: 8

Related Questions