Add support for custom MPI communicator #188

Activity

1. valentinaschueller commented on Apr 10, 2026

   

   valentinaschueller

   on Apr 10, 2026

   More actions

   Hello @fsimonis ! Based on your comment here precice/precice#2544 (comment):

   More as a side note: preCICE is compatible with the Multiple Instruction Multiple Data (MIMD) Model of MPI, where multiple executables are launched in the same world communicator.
   The solver needs to support using a communicator other than world, and the adapter needs to forward the correct communicator to preCICE. None of the official adapters support this, as every participant needs to participate in such a collective split, but it should be easy to implement in in-house codes.
   So, this is possible.

   I wanted to try this with the Python code I work with.
   While the interface suggests that one can pass a communicator of type mpi4py.MPI.Intracomm, it does not work in practice:
   As soon as I pass a communicator, I get a segmentation fault with the following backtrace:

    0 0x0000000000012990 __funlockfile()  :0
    1 0x00000000000b1133 MPI_Comm_rank()  ???:0
    2 0x000000000030f4a9 precice::utils::Parallel::CommState::rank()  :0
    3 0x0000000000281055 precice::impl::ParticipantImpl::ParticipantImpl()  :0
    4 0x0000000000281670 precice::Participant::Participant()  ???:0
    5 0x000000000001f8ec __pyx_pw_9cyprecice_11Participant_3__cinit__()  /tmp/pip-req-build-a0dn74rh/cyprecice/cyprecice.cpp:7020
    6 0x000000000001f8ec __pyx_pw_9cyprecice_11Participant_3__cinit__()  /tmp/pip-req-build-a0dn74rh/cyprecice/cyprecice.cpp:6957
    7 0x0000000000020541 __pyx_tp_new_9cyprecice_Participant()  /tmp/pip-req-build-a0dn74rh/cyprecice/cyprecice.cpp:16785
    8 0x000000000017b65e type_call()  /dev/shm/easybuild/build/Python/3.13.5/GCCcore-14.3.0/Python-3.13.5/Objects/typeobject.c:1985
   

   I have tried MPI.COMM_SELF and MPI.COMM_WORLD; the latter is not technically an Intracomm object but I figured this should behave exactly as passing None.
   I found this issue when I started writing up a reproducible example to ask about it... Does this mean that custom MPI communicators are still not supported for the Python bindings and my error is a result of this?
2. fsimonis commented on Apr 10, 2026

   

   fsimonis

   on Apr 10, 2026

   MemberAuthor

   More actions

   @valentinaschueller Thanks for bringing this up.

   Can you try the fix and report in the PR if this works? #256

   👍React with 👍1Reacted by Valentina Schüller
3. IshaanDesai

   closed this as completedin #256on Apr 10, 2026