jjmaldonis/mpi-parallelization

Examples for MPI Spawning and Splitting, and the differences between two implementations

This repository contains MPI examples using mpi4py.

Examples

Basic Gather Examples

• gather.py uses each core to update a different piece of information in a list and passes that information to all cores using the .gather and .bcast methods. The allgather.py example is similar, but uses the .allgather method instead; it then parses the resulting information to keep only the updated values.

Spawn Multiple Examples

• spawn_multiple_worker.py is a program designed to be run as a worker/child process. spawn_multiple_worker_fortran.f90 is the corresponding FORTRAN example. These programs will be called in the spawning programs below. They distribute a calculation of pi over the cores that were allocated to them and return the reduced-sum value to their parent process.

• spawn.py is the most basic example of spawning, and performs a single spawn. spawn_loop.py has similar functionality, but spawns workers iteratively, waiting for each child process to finish before starting a new one.

• spawn_multiple.py spawns multiple copies of an executable with different data. spawn_multiple_loop.py spawns multiple executables during each iteration of a loop.

• spawn_fortran_multiple.py and spawn_fortran_multiple_loop.py are the analogous programs that call the FORTRAN worker rather than the Python worker.

Split-Spawn Examples

• split_multiple.py is an analogous program to spawn_multiple.py, only rather than using spawn_multiple, it splits the world communicator within the parent process and then spawns processes on the newly created communicators. The worker process can be found in split_multiple_worker.py.

• split_multiple_loop.py splits and spawns multiple executables within a loop.

Notes

This code has only been tested with Open MPI 1.10.2 and mpi4py 2.0.0.

Open MPI sets the environment variable OMPI_MCA_orte_app_num, which is crucial for the spawn_multiple commands. This necessity can be avoided in other MPI implementations by passing the color in to the worker program. This is illustrated in the examples.

Spawn Multiple vs. Split-Spawn

The processes of spawning multiple individuals using Spawn_multiple vs. Split -> Spawn both take about the same amount of time (within the margin of error for the simple tests I ran).

The Spawn Multiple programs can take advantage of the OpenMPI environment variable OMPI_MCA_orte_app_num directly, while the splitting examples inherently cannot. The Spawn Multiple examples can therefore be made to run naturally using MPMD with OpenMPI.

If more MPI APIs were to specify an argument analogous to OMPI_MCA_orte_app_num, the MPI_COMM_Spawn_multiple function would become signficantly more useful.

mpi-parallelization's Issues

License ?

Hello, under what license is this project released under ? I would like to study it to learn from it. Thank you.