Dear all, My MPI application spawns a large number of MPI processes using MPI_Comm_spawn over its total lifetime. Unfortunately, I have experienced that this results in problems for all currently supported OpenMPI versions (2.1, 3.0, 3.1 and 4.0). I have written a short, self-contained program in C (included below) that spawns child processes using MPI_Comm_spawn in an infinite loop, where each child process exits after writing a message to stdout. This short program leads to the following issues: In versions 2.1.2 (Ubuntu package) and 2.1.5 (compiled from source), the program leads to a pipe leak where pipes keep accumulating over time until my MPI application crashes because the maximum number of pipes has been reached. In versions 3.0.3 and 3.1.3 (both compiled from source), there appears to be no pipe leak, but the program crashes with the following error message: PMIX_ERROR: UNREACHABLE in file ptl_tcp_component.c at line 1257
In version 4.0.0 (compiled from source), I have not been able to test this issue very thoroughly because mpiexec ignores the --oversubscribe command-line flag (as detailed in this GitHub issue https://github.com/open-mpi/ompi/issues/6130). This prohibits the oversubscription of processor cores, which means that spawning additional processes immediately results in an error because "not enough slots" are available. A fix for this was proposed recently (https://github.com/open-mpi/ompi/pull/6139), but since the v4.0.x developer branch is being actively developed right now, I decided not go into it. I have found one e-mail thread on this mailing list about a similar problem (https://www.mail-archive.com/users@lists.open-mpi.org/msg10543.html). In this thread, Ralph Castain states that this is a known issue and suggests that it is fixed in the then upcoming v1.3.x release. However, version 1.3 is no longer supported and the issue has reappeared, hence this did not solve the issue. I have created a GitHub gist that contains the output from "ompi_info --all" of all the OpenMPI installations mentioned here, as well as the config.log files for the OpenMPI installations that I compiled from source: https://gist.github.com/ThomasPak/1003160e396bb88dff27e53c53121e0c. I have also attached the code for the short program that demonstrates these issues. For good measure, I have included it directly here as well: """ #include <stdio.h> #include <mpi.h> int main(int argc, char *argv[]) { // Initialize MPI MPI_Init(NULL, NULL); // Get parent MPI_Comm parent; MPI_Comm_get_parent(&parent); // If the process was not spawned if (parent == MPI_COMM_NULL) { puts("I was not spawned!"); // Spawn child process in loop char *cmd = argv[0]; char **cmd_argv = MPI_ARGV_NULL; int maxprocs = 1; MPI_Info info = MPI_INFO_NULL; int root = 0; MPI_Comm comm = MPI_COMM_SELF; MPI_Comm intercomm; int *array_of_errcodes = MPI_ERRCODES_IGNORE; for (;;) { MPI_Comm_spawn(cmd, cmd_argv, maxprocs, info, root, comm, &intercomm, array_of_errcodes); MPI_Comm_disconnect(&intercomm); } // If process was spawned } else { puts("I was spawned!"); MPI_Comm_disconnect(&parent); } // Finalize MPI_Finalize(); } """ Thanks in advance and best wishes, Thomas Pak
mpi-spawn.c
Description: Binary data
_______________________________________________ users mailing list users@lists.open-mpi.org https://lists.open-mpi.org/mailman/listinfo/users
