在mpi编译的fortran可执行文件上从python调用mpirun

Question

I got a fortran simulation code that is only able to run in parallel and has to be compiled with mpi ( make mpi=yes ) on at least 4 cores. I can run the executable, lets name it "test", without problems when i call mpirun -n 4 ./test .

Now I can generate different input files an process the outputs from python. Hence, I want to execute the above command from python to run several simulations. The main problem seems to be, no matter if I use os.system, subprocess.call, .run, .Popen, etc. that for MPI only one process is available (which actually makes sense when python starts one new subprocess):

Eg when using os.system('mpirun -n 4 ./test') or subprocess.run(['mpirun', '-n', '4', './test']) , I get the following output:

starting MPI-3.1 code.
 using    1 nodes with total    1 processors and    1 threads.
 node    0: procs=   1 threads=   1 name=my-pcname

>>>>>>  General information  <<<<<<<
 -------------------------------------------------------------------
 Started              : 22-OCT-2020 18:38:55
 Name of host machine : 
 Current directory    : /home/test
 Compiled on          : Linux (Intel Fortran)
 Compiled             : without OpenMP
 Compiled             : with MPI
 Linked FFT package   : cvecfft_acc         
 Compiled for         :           4  MPI processors
                                  1  OMP threads
 Running on           :           1  MPI processors
                                  1  OMP threads
                                  1  OMP processors per node

>>>> some more information about simulation parameters...

par.f: Mismatch nproc in par.f and MPI nodes:
 Compiled for :            4  MPI processors
 Running on   :            1  MPI processors
 *** STOP *** at location (node            0 ):           3

Interestingly enough I get this output 4 times, which confuses me even more...

Any ideas on how I could get this to work? And sorry, if a similar question was already asked somewhere, I searched for at least an hour and asked colleagues before I decided to post this question here...

I use the Intel fortran compiler ifort 19.0.5.281 20190815 together with OpenMPI 4.0.5

Answer 1

It looks strange. All these messages related to execution environment. Are these produced by your code or is it part of your custom mpirun script?

If I run simple MPI based case

program main

  use mpi

  integer error, id, p

  call MPI_Init ( error )
  call MPI_Comm_size ( MPI_COMM_WORLD, p, error )
  call MPI_Comm_rank ( MPI_COMM_WORLD, id, error )

  write (*,*) 'Hello: ', id, '/', p

  call MPI_Finalize ( error )

end

compiled it with GNU based chain

gnuchain/openmpi-4.0.2-gcc-9.2.0

following way: mpif90 -o hello hello.f90 it works.

And then, if I use simple Python script

import os

os.system('mpirun -np 4 ./hello')

it simply works as expected:

> python ./mpi_run.py
 Hello:            1 /           4
 Hello:            3 /           4
 Hello:            0 /           4
 Hello:            2 /           4

Do you, by any chance, run your Python code itself as an MPI code?