MPI on Linux¶
This document provides the setup and usage of MPI (Message Passing Interface) in OpenMDAO on Linux. We start with installing the necessary packages and test them to make sure they work. Then we will look at a simple example of how to take advantage of MPI in OpenMDAO.
The first package that must be installed is mpi4py which provides Python bindings for MPI. This package requires that your system have an implementation of MPI installed. A quick way to check is to see if you can execute mpirun or mpiexec from your path. You will also need to make sure the MPI C and C++ compiler wrappers mpicc and mpic++ are also in your path. If so, you are ready to install mpi4py with the following command:
pip install mpi4py
The next packages you will want to install are petsc and petsc4py. PETSc stands for “Portable, Extensible Toolkit for Scientific Computation.” It is built on MPI. The package petsc4py is the Python bindings for petsc. To install these packages, first make sure you have a fortran compiler installed, such as the GNU gfortran and make sure it is in your path. Then, run the following command to install both petsc and petsc4py.
pip install --allow-all-external petsc4py
If, for some reason, a pip installation of petsc/petsc4py does not go well, you will need to uninstall your first attempt, and build petsc on your own from source. There can be various pitfalls involved, so it would be wise to consult the most comprehensive guide to installing petsc from source. (including instruction on commonly-used environment variables PETSC_ARCH and PETSC_DIR, which sometimes must be set after installing petsc and before installing petsc4py).
Verify Installed Packages¶
To make sure MPI and petsc are working in your environment, you can use this small petsc4py script:
from petsc4py import PETSc rank = PETSc.COMM_WORLD.getRank() num_ranks = PETSc.COMM_WORLD.getSize() x = PETSc.Vec().createMPI(4) # VecCreateMPI: Creates a parallel vector. size=4 x.setValues([0,1,2,3], [10,20,30,40]) # VecSetValues: Inserts or adds values into certain locations of a vector. x=10, x=20, x=30, x=40 print ('Rank',rank,'has this portion of the MPI vector:', x.getArray() ) # VecGetArray: Returns a pointer to a contiguous array that contains this processor's portion of the vector data. vec_sum = x.sum() # VecSum: Computes the sum of all the components of a vector. 10+20+30+40=100 if rank == 0: print ('Sum of all elements of vector x is',vec_sum,'and was computed using',num_ranks,'MPI processes.')
This script creates a PETSc MPI/parallel vector with four elements, sets the value of those elements, and then computes the total sum of all the elements. You can run the script with two processes using mpirun (or mpiexec):
mpirun -np 2 python petsc_test.py
The output will look something like this:
Rank 1 has this portion of the MPI vector: [ 30. 40.] Rank 0 has this portion of the MPI vector: [ 10. 20.] Sum of all elements of vector x is 100.0 and was computed using 2 MPI processes.
As you can see, because we had a four element vector and two MPI processes, PETSc automatically and evenly divided the vector in half across the two processes. If we tried three processes, PETSc would not be able to split our four element vector up nicely across those processes, yet it would still compute (inefficiently) the correct result:
Rank 1 has this portion of the MPI vector: [ 30.] Rank 2 has this portion of the MPI vector: [ 40.] Rank 0 has this portion of the MPI vector: [ 10. 20.] Sum of all elements of vector x is 100.0 and was computed using 3 MPI processes.