MPI Launch Overview
Overview
Open MPI provides the orterun command to launch MPI applications. However, most people prefer to use one of two aliases: mpirun or mpiexec. For example, to launch 64 MPI tasks named example.x on Carver compute nodes, any of the following commands may be used:
carver% mpirun -np 64 ./example.x
carver% mpiexec -np 64 ./example.x
carver% orterun -np 64 ./example.x
For simplicity, all the examples in this section will use mpirun.
The application launch commands may only be issued from within the batch environment; no MPI applications may be executed on Carver's login nodes.
Options
The Open MPI application launcher provides many runtime options for task placement, debugging, and I/O. Complete descriptions of these options are available on the Open MPI website, or via the command "man mpirun". The following table summarizes a few of the more commonly used options.
| Option | Meaning |
| -v | Produce verbose output. |
| -V | Print version number. |
| -np num | Run num copies of the application. |
| -npernode num | Run num copies of the application on each node. |
| -bynode | Launch processes one per node, cycling by node in a round-robin fashion. |
| -bycore | Launch processes in a block manner, putting all on one node first and then the second (default) |
| -bind-to-core | Bind processes to cores. |
| -stdin rank | Direct stdin to process associated with MPI rank. |
| -tag-output | Tag each line of stdout and stderr with [process_id,rank]. |
| -timestamp-output | Tag each line of stdout and stderr with timestamp. |
Running Multi-Program Multi-Data (MPMD) Applications
Open MPI provides a simple mechanism to launch MPMD applications, using a colon to separate the components on the command line:
carver% mpirun -np 2 ./my_executable1 : -np 2 ./my_executable2
The above example will launch a single parallel application, but the first two processes will be instances of my_executable1, and the second two processes will be instances of my_executable2. In MPI terms, this will be a single MPI_COMM_WORLD, but the my_executable1 processes will be ranks 0 and 1, while the my_executable2 processes will be ranks 2 and 3.
