Error: trace::get_commentry: Invalid communicator index.

[villator]

Hi all, I'm playing with dumpi2otf and the example found at http://mpi.deino.net/mpi_functions/MPI_Intercomm_create.html, which crashes and issues the message "Error: trace::get_commentry: Invalid communicator index."

I'm wondering how to avoid this issue. Maybe this is a bug of libundumpi.

My analysis is the following:

1. "dumpi2otf -f dumpi-2017.11.22.12.28.10.meta -o TRACE.otf" returns the following lines (non-important lines are omitted):

MPI_Comm_split entering at walltime 79071.672523400, cputime 0.019312436 seconds in thread 0.
MPI_Comm oldcomm=2 (MPI_COMM_WORLD)
int color=0
int key=0
MPI_Comm newcomm=4 (user-defined-comm)
MPI_Comm_split returning at walltime 79071.672922008, cputime 0.019510788 seconds in thread 0.

MPI_Intercomm_create entering at walltime 79071.672924778, cputime 0.019513582 seconds in thread 0.
MPI_Comm localcomm=4 (user-defined-comm)
int localleader=0
MPI_Comm remotecomm=2 (MPI_COMM_WORLD)
int remoteleader=1
int tag=52
MPI_Comm newcomm=5 (user-defined-comm)
MPI_Intercomm_create returning at walltime 79071.766963076, cputime 0.064398238 seconds in thread 0.

MPI_Comm_free entering at walltime 79071.767282900, cputime 0.064599472 seconds in thread 0.
MPI_Comm comm=4 (user-defined-comm)
MPI_Comm_free returning at walltime 79071.767291600, cputime 0.064608163 seconds in thread 0.

MPI_Comm_free entering at walltime 79071.767293619, cputime 0.064610199 seconds in thread 0.
MPI_Comm comm=5 (user-defined-comm)
MPI_Comm_free returning at walltime 79071.767298083, cputime 0.064614646 seconds in thread 0.

MPI_Finalize entering at walltime 79071.767299452, cputime 0.064616031 seconds in thread 0.
MPI_Finalize returning at walltime 79071.771893553, cputime 0.066521156 seconds in thread 0.

2. After debugging, I think the problem is the function MPI_Comm_free that does not match the communicators created by MPI_Intercomm_create. dumpi2otf crashes in the function:

int trace::handle_comm_free(const dumpi_comm_free *prm, uint16_t thread,
            const dumpi_time *cpu, const dumpi_time *wall,
            const dumpi_perfinfo *perf, void *userarg) {
        (void) prm;
        (void) thread;
        (void) cpu;
        (void) wall, (void) perf;
        (void) userarg;
        trace *self = (trace*) userarg;
        std::cerr << "DEBUG:  trace::handle_comm_free comm=" << prm->comm << std::endl;
        commentry &the_comm = self->get_commentry(prm->comm); /* the_comm IS NOT FOUND */
        std::cerr << "DEBUG:  success" << std::endl;
        the_comm.freed = wall->stop;
        return 1;
}

trace::commentry& trace::get_commentry(int commhandle) {
        std::pair<commmap_t::iterator, commmap_t::iterator> it =
                this->comms_.equal_range(commhandle);
        if (it.first == it.second) // if no match was found.
            throw "trace::get_commentry:  Invalid communicator index."; /* OUR EXCEPTION */
        commmap_t::iterator ele = it.second;
        --ele;
        return ele->second;
}

3. If we run the command "dumpi2veft -i dumpi-2017.11.22.12.28.10.meta -o TRACE.otf" the output is:
   DEBUG: trace::handle_comm_free comm=4
   DEBUG: success
   DEBUG: trace::handle_comm_free comm=5
   Error: trace::get_commentry: Invalid communicator index.

4. I see that comm 4 is freed, whereas comm=5 is not.

Anyone know what may be happening. Any solution? suggestions?