int ompi_mpi_finalize(void)
{
    int ret;
    static int32_t finalize_has_already_started = 0;
    opal_list_item_t *item;
    ompi_proc_t** procs;
    size_t nprocs;
    OPAL_TIMING_DECLARE(tm);
    OPAL_TIMING_INIT_EXT(&tm, OPAL_TIMING_GET_TIME_OF_DAY);


    /* Be a bit social if an erroneous program calls MPI_FINALIZE in
       two different threads, otherwise we may deadlock in
       ompi_comm_free() (or run into other nasty lions, tigers, or
       bears) */

    if (! opal_atomic_cmpset_32(&finalize_has_already_started, 0, 1)) {
        /* Note that if we're already finalized, we cannot raise an
           MPI exception.  The best that we can do is write something
           to stderr. */
        char hostname[MAXHOSTNAMELEN];
        pid_t pid = getpid();
        gethostname(hostname, sizeof(hostname));

        opal_show_help("help-mpi-runtime.txt",
                       "mpi_finalize:invoked_multiple_times",
                       true, hostname, pid);
        return MPI_ERR_OTHER;
    }

    ompi_mpiext_fini();

    /* Per MPI-2:4.8, we have to free MPI_COMM_SELF before doing
       anything else in MPI_FINALIZE (to include setting up such that
       MPI_FINALIZED will return true). */

    if (NULL != ompi_mpi_comm_self.comm.c_keyhash) {
        ompi_attr_delete_all(COMM_ATTR, &ompi_mpi_comm_self,
                             ompi_mpi_comm_self.comm.c_keyhash);
        OBJ_RELEASE(ompi_mpi_comm_self.comm.c_keyhash);
        ompi_mpi_comm_self.comm.c_keyhash = NULL;
    }

    /* Proceed with MPI_FINALIZE */

    ompi_mpi_finalized = true;

    /* As finalize is the last legal MPI call, we are allowed to force the release
     * of the user buffer used for bsend, before going anywhere further.
     */
    (void)mca_pml_base_bsend_detach(NULL, NULL);

#if OPAL_ENABLE_PROGRESS_THREADS == 0
    opal_progress_set_event_flag(OPAL_EVLOOP_ONCE | OPAL_EVLOOP_NONBLOCK);
#endif

    /* Redo ORTE calling opal_progress_event_users_increment() during
       MPI lifetime, to get better latency when not using TCP */
    opal_progress_event_users_increment();

    /* check to see if we want timing information */
    OPAL_TIMING_MSTART((&tm,"time to execute finalize barrier"));

    /* NOTE: MPI-2.1 requires that MPI_FINALIZE is "collective" across
       *all* connected processes.  This only means that all processes
       have to call it.  It does *not* mean that all connected
       processes need to synchronize (either directly or indirectly).

       For example, it is quite easy to construct complicated
       scenarios where one job is "connected" to another job via
       transitivity, but have no direct knowledge of each other.
       Consider the following case: job A spawns job B, and job B
       later spawns job C.  A "connectedness" graph looks something
       like this:

           A <--> B <--> C

       So what are we *supposed* to do in this case?  If job A is
       still connected to B when it calls FINALIZE, should it block
       until jobs B and C also call FINALIZE?

       After lengthy discussions many times over the course of this
       project, the issue was finally decided at the Louisville Feb
       2009 meeting: no.

       Rationale:

       - "Collective" does not mean synchronizing.  It only means that
         every process call it.  Hence, in this scenario, every
         process in A, B, and C must call FINALIZE.

       - KEY POINT: if A calls FINALIZE, then it is erroneous for B or
         C to try to communicate with A again.

       - Hence, OMPI is *correct* to only effect a barrier across each
         jobs' MPI_COMM_WORLD before exiting.  Specifically, if A
         calls FINALIZE long before B or C, it's *correct* if A exits
         at any time (and doesn't notify B or C that it is exiting).

       - Arguably, if B or C do try to communicate with the now-gone
         A, OMPI should try to print a nice error ("you tried to
         communicate with a job that is already gone...") instead of
         segv or other Badness.  However, that is an *extremely*
         difficult problem -- sure, it's easy for A to tell B that it
         is finalizing, but how can A tell C?  A doesn't even know
         about C.  You'd need to construct a "connected" graph in a
         distributed fashion, which is fraught with race conditions,
         etc.

      Hence, our conclusion is: OMPI is *correct* in its current
      behavior (of only doing a barrier across its own COMM_WORLD)
      before exiting.  Any problems that occur are as a result of
      erroneous MPI applications.  We *could* tighten up the erroneous
      cases and ensure that we print nice error messages / don't
      crash, but that is such a difficult problem that we decided we
      have many other, much higher priority issues to handle that deal
      with non-erroneous cases. */

    /* Wait for everyone to reach this point.  This is a grpcomm
       barrier instead of an MPI barrier for (at least) two reasons:

       1. An MPI barrier doesn't ensure that all messages have been
          transmitted before exiting (e.g., a BTL can lie and buffer a
          message without actually injecting it to the network, and
          therefore require further calls to that BTL's progress), so
          the possibility of a stranded message exists.

       2. If the MPI communication is using an unreliable transport,
          there's a problem of knowing that everyone has *left* the
          barrier.  E.g., one proc can send its ACK to the barrier
          message to a peer and then leave the barrier, but the ACK
          can get lost and therefore the peer is left in the barrier.

       Point #1 has been known for a long time; point #2 emerged after
       we added the first unreliable BTL to Open MPI and fixed the
       del_procs behavior around May of 2014 (see
       https://svn.open-mpi.org/trac/ompi/ticket/4669#comment:4 for
       more details). */
    opal_pmix.fence(NULL, 0);

    /* check for timing request - get stop time and report elapsed
     time if so */
    OPAL_TIMING_MSTOP(&tm);
    OPAL_TIMING_DELTAS(ompi_enable_timing, &tm);
    OPAL_TIMING_REPORT(ompi_enable_timing_ext, &tm);
    OPAL_TIMING_RELEASE(&tm);

    /*
     * Shutdown the Checkpoint/Restart Mech.
     */
    if (OMPI_SUCCESS != (ret = ompi_cr_finalize())) {
        OMPI_ERROR_LOG(ret);
    }

    /* Shut down any bindings-specific issues: C++, F77, F90 */

    /* Remove all memory associated by MPI_REGISTER_DATAREP (per
       MPI-2:9.5.3, there is no way for an MPI application to
       *un*register datareps, but we don't want the OMPI layer causing
       memory leaks). */
    while (NULL != (item = opal_list_remove_first(&ompi_registered_datareps))) {
        OBJ_RELEASE(item);
    }
    OBJ_DESTRUCT(&ompi_registered_datareps);

    /* Remove all F90 types from the hash tables. As the OBJ_DESTRUCT will
     * call a special destructor able to release predefined types, we can
     * simply call the OBJ_DESTRUCT on the hash table and all memory will
     * be correctly released.
     */
    OBJ_DESTRUCT( &ompi_mpi_f90_integer_hashtable );
    OBJ_DESTRUCT( &ompi_mpi_f90_real_hashtable );
    OBJ_DESTRUCT( &ompi_mpi_f90_complex_hashtable );

    /* Free communication objects */

    /* free file resources */
    if (OMPI_SUCCESS != (ret = ompi_file_finalize())) {
        return ret;
    }

    /* free window resources */
    if (OMPI_SUCCESS != (ret = ompi_win_finalize())) {
        return ret;
    }
    if (OMPI_SUCCESS != (ret = ompi_osc_base_finalize())) {
        return ret;
    }

    /* free communicator resources. this MUST come before finalizing the PML
     * as this will call into the pml */
    if (OMPI_SUCCESS != (ret = ompi_comm_finalize())) {
        return ret;
    }

    nprocs = 0;
    procs = ompi_proc_world(&nprocs);
    MCA_PML_CALL(del_procs(procs, nprocs));
    free(procs);

    /* free pml resource */
    if(OMPI_SUCCESS != (ret = mca_pml_base_finalize())) {
      return ret;
    }

    /* free requests */
    if (OMPI_SUCCESS != (ret = ompi_request_finalize())) {
        return ret;
    }

    if (OMPI_SUCCESS != (ret = ompi_message_finalize())) {
        return ret;
    }

    /* If requested, print out a list of memory allocated by ALLOC_MEM
       but not freed by FREE_MEM */
    if (0 != ompi_debug_show_mpi_alloc_mem_leaks) {
        mca_mpool_base_tree_print(ompi_debug_show_mpi_alloc_mem_leaks);
    }

    /* Now that all MPI objects dealing with communications are gone,
       shut down MCA types having to do with communications */
    if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_pml_base_framework) ) ) {
        OMPI_ERROR_LOG(ret);
        return ret;
    }

    /* shut down buffered send code */
    mca_pml_base_bsend_fini();

#if OPAL_ENABLE_FT_CR == 1
    /*
     * Shutdown the CRCP Framework, must happen after PML shutdown
     */
    if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_crcp_base_framework) ) ) {
        OMPI_ERROR_LOG(ret);
        return ret;
    }
#endif

    /* Free secondary resources */

    /* free attr resources */
    if (OMPI_SUCCESS != (ret = ompi_attr_finalize())) {
        return ret;
    }

    /* free group resources */
    if (OMPI_SUCCESS != (ret = ompi_group_finalize())) {
        return ret;
    }

    /* finalize the pubsub functions */
    if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_pubsub_base_framework) ) ) {
        return ret;
    }

    /* finalize the DPM framework */
    if ( OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_dpm_base_framework))) {
        return ret;
    }

    /* free internal error resources */
    if (OMPI_SUCCESS != (ret = ompi_errcode_intern_finalize())) {
        return ret;
    }

    /* free error code resources */
    if (OMPI_SUCCESS != (ret = ompi_mpi_errcode_finalize())) {
        return ret;
    }

    /* free errhandler resources */
    if (OMPI_SUCCESS != (ret = ompi_errhandler_finalize())) {
        return ret;
    }

    /* Free all other resources */

    /* free op resources */
    if (OMPI_SUCCESS != (ret = ompi_op_finalize())) {
        return ret;
    }

    /* free ddt resources */
    if (OMPI_SUCCESS != (ret = ompi_datatype_finalize())) {
        return ret;
    }

    /* free info resources */
    if (OMPI_SUCCESS != (ret = ompi_info_finalize())) {
        return ret;
    }

    /* Close down MCA modules */

    /* io is opened lazily, so it's only necessary to close it if it
       was actually opened */
    if (0 < ompi_io_base_framework.framework_refcnt) {
        /* May have been "opened" multiple times. We want it closed now */
        ompi_io_base_framework.framework_refcnt = 1;

        if (OMPI_SUCCESS != mca_base_framework_close(&ompi_io_base_framework)) {
            return ret;
        }
    }
    (void) mca_base_framework_close(&ompi_topo_base_framework);
    if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_osc_base_framework))) {
        return ret;
    }
    if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_coll_base_framework))) {
        return ret;
    }
    if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_bml_base_framework))) {
        return ret;
    }
    if (OMPI_SUCCESS != (ret = mca_base_framework_close(&opal_mpool_base_framework))) {
        return ret;
    }
    if (OMPI_SUCCESS != (ret = mca_base_framework_close(&opal_rcache_base_framework))) {
        return ret;
    }
    if (OMPI_SUCCESS != (ret = mca_base_framework_close(&opal_allocator_base_framework))) {
        return ret;
    }

    /* free proc resources */
    if ( OMPI_SUCCESS != (ret = ompi_proc_finalize())) {
        return ret;
    }

    if (NULL != ompi_mpi_main_thread) {
        OBJ_RELEASE(ompi_mpi_main_thread);
        ompi_mpi_main_thread = NULL;
    }

    /* Leave the RTE */

    if (OMPI_SUCCESS != (ret = ompi_rte_finalize())) {
        return ret;
    }
    ompi_rte_initialized = false;

    /* now close the rte framework */
    if (OMPI_SUCCESS != (ret = mca_base_framework_close(&ompi_rte_base_framework) ) ) {
        OMPI_ERROR_LOG(ret);
        return ret;
    }

    if (OPAL_SUCCESS != (ret = opal_finalize_util())) {
        return ret;
    }

    /* All done */

    return MPI_SUCCESS;
}
Exemple #2
0
int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
{
    int ret;
    ompi_proc_t** procs;
    size_t nprocs;
    char *error = NULL;
    char *cmd=NULL, *av=NULL;
    ompi_errhandler_errtrk_t errtrk;
    OPAL_TIMING_DECLARE(tm);
    OPAL_TIMING_INIT_EXT(&tm, OPAL_TIMING_GET_TIME_OF_DAY);

    /* bitflag of the thread level support provided. To be used
     * for the modex in order to work in heterogeneous environments. */
    uint8_t threadlevel_bf;

    /* Ensure that we were not already initialized or finalized.

       This lock is held for the duration of ompi_mpi_init() and
       ompi_mpi_finalize().  Hence, if we get it, then no other thread
       is inside the critical section (and we don't have to check the
       *_started bool variables). */
    opal_mutex_lock(&ompi_mpi_bootstrap_mutex);
    if (ompi_mpi_finalized) {
        opal_show_help("help-mpi-runtime.txt",
                       "mpi_init: already finalized", true);
        opal_mutex_unlock(&ompi_mpi_bootstrap_mutex);
        return MPI_ERR_OTHER;
    } else if (ompi_mpi_initialized) {
        opal_show_help("help-mpi-runtime.txt",
                       "mpi_init: invoked multiple times", true);
        opal_mutex_unlock(&ompi_mpi_bootstrap_mutex);
        return MPI_ERR_OTHER;
    }

    /* Indicate that we have *started* MPI_INIT* */
    ompi_mpi_init_started = true;

    /* Setup enough to check get/set MCA params */

    if (OPAL_SUCCESS != (ret = opal_init_util(&argc, &argv))) {
        error = "ompi_mpi_init: opal_init_util failed";
        goto error;
    }

    /* Convince OPAL to use our naming scheme */
    opal_process_name_print = _process_name_print_for_opal;
    opal_compare_proc = _process_name_compare;
    opal_convert_string_to_process_name = _convert_string_to_process_name;
    opal_convert_process_name_to_string = _convert_process_name_to_string;
    opal_proc_for_name = ompi_proc_for_name;

    /* Register MCA variables */
    if (OPAL_SUCCESS != (ret = ompi_register_mca_variables())) {
        error = "ompi_mpi_init: ompi_register_mca_variables failed";
        goto error;
    }

    if (OPAL_SUCCESS != (ret = opal_arch_set_fortran_logical_size(sizeof(ompi_fortran_logical_t)))) {
        error = "ompi_mpi_init: opal_arch_set_fortran_logical_size failed";
        goto error;
    }

    /* _After_ opal_init_util() but _before_ orte_init(), we need to
       set an MCA param that tells libevent that it's ok to use any
       mechanism in libevent that is available on this platform (e.g.,
       epoll and friends).  Per opal/event/event.s, we default to
       select/poll -- but we know that MPI processes won't be using
       pty's with the event engine, so it's ok to relax this
       constraint and let any fd-monitoring mechanism be used. */

    ret = mca_base_var_find("opal", "event", "*", "event_include");
    if (ret >= 0) {
        char *allvalue = "all";
        /* We have to explicitly "set" the MCA param value here
           because libevent initialization will re-register the MCA
           param and therefore override the default. Setting the value
           here puts the desired value ("all") in different storage
           that is not overwritten if/when the MCA param is
           re-registered. This is unless the user has specified a different
           value for this MCA parameter. Make sure we check to see if the
           default is specified before forcing "all" in case that is not what
           the user desires. Note that we do *NOT* set this value as an
           environment variable, just so that it won't be inherited by
           any spawned processes and potentially cause unintented
           side-effects with launching RTE tools... */
        mca_base_var_set_value(ret, allvalue, 4, MCA_BASE_VAR_SOURCE_DEFAULT, NULL);
    }

    OPAL_TIMING_MSTART((&tm,"time from start to completion of rte_init"));

    /* if we were not externally started, then we need to setup
     * some envars so the MPI_INFO_ENV can get the cmd name
     * and argv (but only if the user supplied a non-NULL argv!), and
     * the requested thread level
     */
    if (NULL == getenv("OMPI_COMMAND") && NULL != argv && NULL != argv[0]) {
        asprintf(&cmd, "OMPI_COMMAND=%s", argv[0]);
        putenv(cmd);
    }
    if (NULL == getenv("OMPI_ARGV") && 1 < argc) {
        char *tmp;
        tmp = opal_argv_join(&argv[1], ' ');
        asprintf(&av, "OMPI_ARGV=%s", tmp);
        free(tmp);
        putenv(av);
    }

    /* open the rte framework */
    if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_rte_base_framework, 0))) {
        error = "ompi_rte_base_open() failed";
        goto error;
    }
    /* no select is required as this is a static framework */

    /* Setup RTE */
    if (OMPI_SUCCESS != (ret = ompi_rte_init(NULL, NULL))) {
        error = "ompi_mpi_init: ompi_rte_init failed";
        goto error;
    }
    ompi_rte_initialized = true;

    /* check for timing request - get stop time and report elapsed time if so */
    OPAL_TIMING_MNEXT((&tm,"time from completion of rte_init to modex"));

    /* if hwloc is available but didn't get setup for some
     * reason, do so now
     */
    if (NULL == opal_hwloc_topology) {
        if (OPAL_SUCCESS != (ret = opal_hwloc_base_get_topology())) {
            error = "Topology init";
            goto error;
        }
    }

    /* Register the default errhandler callback  */
    errtrk.status = OPAL_ERROR;
    errtrk.active = true;
    opal_pmix.register_errhandler(NULL, ompi_errhandler_callback,
                                  ompi_errhandler_registration_callback,
                                  (void*)&errtrk);
    OMPI_WAIT_FOR_COMPLETION(errtrk.active);
    if (OPAL_SUCCESS != errtrk.status) {
        error = "Error handler registration";
        ret = errtrk.status;
        goto error;
    }

    /* Figure out the final MPI thread levels.  If we were not
       compiled for support for MPI threads, then don't allow
       MPI_THREAD_MULTIPLE.  Set this stuff up here early in the
       process so that other components can make decisions based on
       this value. */

    ompi_mpi_thread_level(requested, provided);

    /* determine the bitflag belonging to the threadlevel_support provided */
    memset ( &threadlevel_bf, 0, sizeof(uint8_t));
    OMPI_THREADLEVEL_SET_BITFLAG ( ompi_mpi_thread_provided, threadlevel_bf );

#if OMPI_ENABLE_THREAD_MULTIPLE
    /* add this bitflag to the modex */
    OPAL_MODEX_SEND_STRING(ret, OPAL_PMIX_GLOBAL,
                           "MPI_THREAD_LEVEL", &threadlevel_bf, sizeof(uint8_t));
    if (OPAL_SUCCESS != ret) {
        error = "ompi_mpi_init: modex send thread level";
        goto error;
    }
#endif

    /* If thread support was enabled, then setup OPAL to allow for
       them. */
    if ((OPAL_ENABLE_PROGRESS_THREADS == 1) ||
            (*provided != MPI_THREAD_SINGLE)) {
        opal_set_using_threads(true);
    }

    /* initialize datatypes. This step should be done early as it will
     * create the local convertor and local arch used in the proc
     * init.
     */
    if (OMPI_SUCCESS != (ret = ompi_datatype_init())) {
        error = "ompi_datatype_init() failed";
        goto error;
    }

    /* Initialize OMPI procs */
    if (OMPI_SUCCESS != (ret = ompi_proc_init())) {
        error = "mca_proc_init() failed";
        goto error;
    }

    /* Initialize the op framework. This has to be done *after*
       ddt_init, but befor mca_coll_base_open, since some collective
       modules (e.g., the hierarchical coll component) may need ops in
       their query function. */
    if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_op_base_framework, 0))) {
        error = "ompi_op_base_open() failed";
        goto error;
    }
    if (OMPI_SUCCESS !=
            (ret = ompi_op_base_find_available(OPAL_ENABLE_PROGRESS_THREADS,
                    ompi_mpi_thread_multiple))) {
        error = "ompi_op_base_find_available() failed";
        goto error;
    }
    if (OMPI_SUCCESS != (ret = ompi_op_init())) {
        error = "ompi_op_init() failed";
        goto error;
    }

    /* Open up MPI-related MCA components */

    if (OMPI_SUCCESS != (ret = mca_base_framework_open(&opal_allocator_base_framework, 0))) {
        error = "mca_allocator_base_open() failed";
        goto error;
    }
    if (OMPI_SUCCESS != (ret = mca_base_framework_open(&opal_rcache_base_framework, 0))) {
        error = "mca_rcache_base_open() failed";
        goto error;
    }
    if (OMPI_SUCCESS != (ret = mca_base_framework_open(&opal_mpool_base_framework, 0))) {
        error = "mca_mpool_base_open() failed";
        goto error;
    }
    if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_bml_base_framework, 0))) {
        error = "mca_bml_base_open() failed";
        goto error;
    }
    if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_pml_base_framework, 0))) {
        error = "mca_pml_base_open() failed";
        goto error;
    }
    if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_coll_base_framework, 0))) {
        error = "mca_coll_base_open() failed";
        goto error;
    }

    if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_osc_base_framework, 0))) {
        error = "ompi_osc_base_open() failed";
        goto error;
    }

#if OPAL_ENABLE_FT_CR == 1
    if (OMPI_SUCCESS != (ret = mca_base_framework_open(&ompi_crcp_base_framework, 0))) {
        error = "ompi_crcp_base_open() failed";
        goto error;
    }
#endif

    /* In order to reduce the common case for MPI apps (where they
       don't use MPI-2 IO or MPI-1 topology functions), the io and
       topo frameworks are initialized lazily, at the first use of
       relevant functions (e.g., MPI_FILE_*, MPI_CART_*, MPI_GRAPH_*),
       so they are not opened here. */

    /* Select which MPI components to use */

    if (OMPI_SUCCESS !=
            (ret = mca_mpool_base_init(OPAL_ENABLE_PROGRESS_THREADS,
                                       ompi_mpi_thread_multiple))) {
        error = "mca_mpool_base_init() failed";
        goto error;
    }

    if (OMPI_SUCCESS !=
            (ret = mca_pml_base_select(OPAL_ENABLE_PROGRESS_THREADS,
                                       ompi_mpi_thread_multiple))) {
        error = "mca_pml_base_select() failed";
        goto error;
    }

    /* check for timing request - get stop time and report elapsed time if so */
    OPAL_TIMING_MNEXT((&tm,"time to execute modex"));

    /* exchange connection info - this function may also act as a barrier
     * if data exchange is required. The modex occurs solely across procs
     * in our job. If a barrier is required, the "modex" function will
     * perform it internally */
    OPAL_MODEX();

    OPAL_TIMING_MNEXT((&tm,"time from modex to first barrier"));

    /* select buffered send allocator component to be used */
    if( OMPI_SUCCESS !=
            (ret = mca_pml_base_bsend_init(ompi_mpi_thread_multiple))) {
        error = "mca_pml_base_bsend_init() failed";
        goto error;
    }

    if (OMPI_SUCCESS !=
            (ret = mca_coll_base_find_available(OPAL_ENABLE_PROGRESS_THREADS,
                    ompi_mpi_thread_multiple))) {
        error = "mca_coll_base_find_available() failed";
        goto error;
    }

    if (OMPI_SUCCESS !=
            (ret = ompi_osc_base_find_available(OPAL_ENABLE_PROGRESS_THREADS,
                    ompi_mpi_thread_multiple))) {
        error = "ompi_osc_base_find_available() failed";
        goto error;
    }

#if OPAL_ENABLE_FT_CR == 1
    if (OMPI_SUCCESS != (ret = ompi_crcp_base_select() ) ) {
        error = "ompi_crcp_base_select() failed";
        goto error;
    }
#endif

    /* io and topo components are not selected here -- see comment
       above about the io and topo frameworks being loaded lazily */

    /* Initialize each MPI handle subsystem */
    /* initialize requests */
    if (OMPI_SUCCESS != (ret = ompi_request_init())) {
        error = "ompi_request_init() failed";
        goto error;
    }

    if (OMPI_SUCCESS != (ret = ompi_message_init())) {
        error = "ompi_message_init() failed";
        goto error;
    }

    /* initialize info */
    if (OMPI_SUCCESS != (ret = ompi_mpiinfo_init())) {
        error = "ompi_info_init() failed";
        goto error;
    }

    /* initialize error handlers */
    if (OMPI_SUCCESS != (ret = ompi_errhandler_init())) {
        error = "ompi_errhandler_init() failed";
        goto error;
    }

    /* initialize error codes */
    if (OMPI_SUCCESS != (ret = ompi_mpi_errcode_init())) {
        error = "ompi_mpi_errcode_init() failed";
        goto error;
    }

    /* initialize internal error codes */
    if (OMPI_SUCCESS != (ret = ompi_errcode_intern_init())) {
        error = "ompi_errcode_intern_init() failed";
        goto error;
    }

    /* initialize groups  */
    if (OMPI_SUCCESS != (ret = ompi_group_init())) {
        error = "ompi_group_init() failed";
        goto error;
    }

    /* initialize communicators */
    if (OMPI_SUCCESS != (ret = ompi_comm_init())) {
        error = "ompi_comm_init() failed";
        goto error;
    }

    /* initialize file handles */
    if (OMPI_SUCCESS != (ret = ompi_file_init())) {
        error = "ompi_file_init() failed";
        goto error;
    }

    /* initialize windows */
    if (OMPI_SUCCESS != (ret = ompi_win_init())) {
        error = "ompi_win_init() failed";
        goto error;
    }

    /* initialize attribute meta-data structure for comm/win/dtype */
    if (OMPI_SUCCESS != (ret = ompi_attr_init())) {
        error = "ompi_attr_init() failed";
        goto error;
    }

    /* identify the architectures of remote procs and setup
     * their datatype convertors, if required
     */
    if (OMPI_SUCCESS != (ret = ompi_proc_complete_init())) {
        error = "ompi_proc_complete_init failed";
        goto error;
    }

    /* start PML/BTL's */
    ret = MCA_PML_CALL(enable(true));
    if( OMPI_SUCCESS != ret ) {
        error = "PML control failed";
        goto error;
    }

    /* some btls/mtls require we call add_procs with all procs in the job.
     * since the btls/mtls have no visibility here it is up to the pml to
     * convey this requirement */
    if (mca_pml_base_requires_world ()) {
        if (NULL == (procs = ompi_proc_world (&nprocs))) {
            error = "ompi_proc_get_allocated () failed";
            goto error;
        }
    } else {
        /* add all allocated ompi_proc_t's to PML (below the add_procs limit this
         * behaves identically to ompi_proc_world ()) */
        if (NULL == (procs = ompi_proc_get_allocated (&nprocs))) {
            error = "ompi_proc_get_allocated () failed";
            goto error;
        }
    }
    ret = MCA_PML_CALL(add_procs(procs, nprocs));
    free(procs);
    /* If we got "unreachable", then print a specific error message.
       Otherwise, if we got some other failure, fall through to print
       a generic message. */
    if (OMPI_ERR_UNREACH == ret) {
        opal_show_help("help-mpi-runtime.txt",
                       "mpi_init:startup:pml-add-procs-fail", true);
        error = NULL;
        goto error;
    } else if (OMPI_SUCCESS != ret) {
        error = "PML add procs failed";
        goto error;
    }

    MCA_PML_CALL(add_comm(&ompi_mpi_comm_world.comm));
    MCA_PML_CALL(add_comm(&ompi_mpi_comm_self.comm));

    /*
     * Dump all MCA parameters if requested
     */
    if (ompi_mpi_show_mca_params) {
        ompi_show_all_mca_params(ompi_mpi_comm_world.comm.c_my_rank,
                                 nprocs,
                                 ompi_process_info.nodename);
    }

    /* Do we need to wait for a debugger? */
    ompi_rte_wait_for_debugger();

    /* Next timing measurement */
    OPAL_TIMING_MNEXT((&tm,"time to execute barrier"));

    /* wait for everyone to reach this point - this is a hard
     * barrier requirement at this time, though we hope to relax
     * it at a later point */
    opal_pmix.fence(NULL, 0);

    /* check for timing request - get stop time and report elapsed
       time if so, then start the clock again */
    OPAL_TIMING_MNEXT((&tm,"time from barrier to complete mpi_init"));

#if OPAL_ENABLE_PROGRESS_THREADS == 0
    /* Start setting up the event engine for MPI operations.  Don't
       block in the event library, so that communications don't take
       forever between procs in the dynamic code.  This will increase
       CPU utilization for the remainder of MPI_INIT when we are
       blocking on RTE-level events, but may greatly reduce non-TCP
       latency. */
    opal_progress_set_event_flag(OPAL_EVLOOP_NONBLOCK);
#endif

    /* wire up the mpi interface, if requested.  Do this after the
       non-block switch for non-TCP performance.  Do before the
       polling change as anyone with a complex wire-up is going to be
       using the oob. */
    if (OMPI_SUCCESS != (ret = ompi_init_preconnect_mpi())) {
        error = "ompi_mpi_do_preconnect_all() failed";
        goto error;
    }

    /* Setup the dynamic process management (DPM) subsystem */
    if (OMPI_SUCCESS != (ret = ompi_dpm_init())) {
        error = "ompi_dpm_init() failed";
        goto error;
    }

    /* Determine the overall threadlevel support of all processes
       in MPI_COMM_WORLD. This has to be done before calling
       coll_base_comm_select, since some of the collective components
       e.g. hierarch, might create subcommunicators. The threadlevel
       requested by all processes is required in order to know
       which cid allocation algorithm can be used. */
    if ( OMPI_SUCCESS !=
            ( ret = ompi_comm_cid_init ())) {
        error = "ompi_mpi_init: ompi_comm_cid_init failed";
        goto error;
    }

    /* Init coll for the comms. This has to be after dpm_base_select,
       (since dpm.mark_dyncomm is not set in the communicator creation
       function else), but before dpm.dyncom_init, since this function
       might require collective for the CID allocation. */
    if (OMPI_SUCCESS !=
            (ret = mca_coll_base_comm_select(MPI_COMM_WORLD))) {
        error = "mca_coll_base_comm_select(MPI_COMM_WORLD) failed";
        goto error;
    }

    if (OMPI_SUCCESS !=
            (ret = mca_coll_base_comm_select(MPI_COMM_SELF))) {
        error = "mca_coll_base_comm_select(MPI_COMM_SELF) failed";
        goto error;
    }

    /* Check whether we have been spawned or not.  We introduce that
       at the very end, since we need collectives, datatypes, ptls
       etc. up and running here.... */
    if (OMPI_SUCCESS != (ret = ompi_dpm_dyn_init())) {
        error = "ompi_dpm_dyn_init() failed";
        goto error;
    }

    /*
     * Startup the Checkpoint/Restart Mech.
     * Note: Always do this so tools don't hang when
     * in a non-checkpointable build
     */
    if (OMPI_SUCCESS != (ret = ompi_cr_init())) {
        error = "ompi_cr_init";
        goto error;
    }

    /* Undo OPAL calling opal_progress_event_users_increment() during
       opal_init, to get better latency when not using TCP.  Do
       this *after* dyn_init, as dyn init uses lots of RTE
       communication and we don't want to hinder the performance of
       that code. */
    opal_progress_event_users_decrement();

    /* see if yield_when_idle was specified - if so, use it */
    opal_progress_set_yield_when_idle(ompi_mpi_yield_when_idle);

    /* negative value means use default - just don't do anything */
    if (ompi_mpi_event_tick_rate >= 0) {
        opal_progress_set_event_poll_rate(ompi_mpi_event_tick_rate);
    }

    /* At this point, we are fully configured and in MPI mode.  Any
       communication calls here will work exactly like they would in
       the user's code.  Setup the connections between procs and warm
       them up with simple sends, if requested */

    if (OMPI_SUCCESS != (ret = ompi_mpiext_init())) {
        error = "ompi_mpiext_init";
        goto error;
    }

    /* Fall through */
error:
    if (ret != OMPI_SUCCESS) {
        /* Only print a message if one was not already printed */
        if (NULL != error) {
            const char *err_msg = opal_strerror(ret);
            opal_show_help("help-mpi-runtime.txt",
                           "mpi_init:startup:internal-failure", true,
                           "MPI_INIT", "MPI_INIT", error, err_msg, ret);
        }
        opal_mutex_unlock(&ompi_mpi_bootstrap_mutex);
        return ret;
    }

    /* Initialize the registered datarep list to be empty */
    OBJ_CONSTRUCT(&ompi_registered_datareps, opal_list_t);

    /* Initialize the arrays used to store the F90 types returned by the
     *  MPI_Type_create_f90_XXX functions.
     */
    OBJ_CONSTRUCT( &ompi_mpi_f90_integer_hashtable, opal_hash_table_t);
    opal_hash_table_init(&ompi_mpi_f90_integer_hashtable, 16 /* why not? */);

    OBJ_CONSTRUCT( &ompi_mpi_f90_real_hashtable, opal_hash_table_t);
    opal_hash_table_init(&ompi_mpi_f90_real_hashtable, FLT_MAX_10_EXP);

    OBJ_CONSTRUCT( &ompi_mpi_f90_complex_hashtable, opal_hash_table_t);
    opal_hash_table_init(&ompi_mpi_f90_complex_hashtable, FLT_MAX_10_EXP);

    /* All done.  Wasn't that simple? */

    ompi_mpi_initialized = true;

    /* Finish last measurement, output results
     * and clear timing structure */
    OPAL_TIMING_MSTOP(&tm);
    OPAL_TIMING_DELTAS(ompi_enable_timing, &tm);
    OPAL_TIMING_REPORT(ompi_enable_timing_ext, &tm);
    OPAL_TIMING_RELEASE(&tm);

    opal_mutex_unlock(&ompi_mpi_bootstrap_mutex);
    return MPI_SUCCESS;
}