ompi/mpi_init: fix barrier

Relax CPU usage pressure from the application processes when doing
modex and barrier in ompi_mpi_init.

We see significant latencies in SLURM/pmix plugin barrier progress
because app processes are aggressively call opal_progress pushing
away daemon process doing collective progress.

Author: artpol84

ompi/runtime/ompi_mpi_init.c

@@ -280,6 +280,7 @@ opal_list_t ompi_registered_datareps = {{0}};
 
 bool ompi_enable_timing = false, ompi_enable_timing_ext = false;
 extern bool ompi_mpi_yield_when_idle;
+extern bool ompi_mpi_lazy_wait_in_init;
 extern int ompi_mpi_event_tick_rate;
 
 /**
@@ -532,7 +533,12 @@ int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
     opal_pmix.register_evhandler(NULL, &info, ompi_errhandler_callback,
                                  ompi_errhandler_registration_callback,
                                  (void*)&errtrk);
-    OMPI_WAIT_FOR_COMPLETION(errtrk.active);
+    if( ompi_mpi_lazy_wait_in_init ){
+        OMPI_LAZY_WAIT_FOR_COMPLETION(errtrk.active);
+    } else {
+        OMPI_WAIT_FOR_COMPLETION(errtrk.active);
+    }
+
     OPAL_LIST_DESTRUCT(&info);
     if (OPAL_SUCCESS != errtrk.status) {
         error = "Error handler registration";
@@ -658,7 +664,11 @@ int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
         if (NULL != opal_pmix.fence_nb) {
             opal_pmix.fence_nb(NULL, opal_pmix_collect_all_data,
                                fence_release, (void*)&active);
-            OMPI_WAIT_FOR_COMPLETION(active);
+            if( ompi_mpi_lazy_wait_in_init ){
+                OMPI_LAZY_WAIT_FOR_COMPLETION(active);
+            } else {
+                OMPI_WAIT_FOR_COMPLETION(active);
+            }
         } else {
             opal_pmix.fence(NULL, opal_pmix_collect_all_data);
         }
@@ -835,7 +845,11 @@ int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
         if (NULL != opal_pmix.fence_nb) {
             opal_pmix.fence_nb(NULL, opal_pmix_collect_all_data,
                                fence_release, (void*)&active);
-            OMPI_WAIT_FOR_COMPLETION(active);
+            if( ompi_mpi_lazy_wait_in_init ){
+                OMPI_LAZY_WAIT_FOR_COMPLETION(active);
+            } else {
+                OMPI_WAIT_FOR_COMPLETION(active);
+            }
         } else {
             opal_pmix.fence(NULL, opal_pmix_collect_all_data);
         }

ompi/runtime/ompi_mpi_params.c

@@ -60,6 +60,7 @@ bool ompi_have_sparse_group_storage = OPAL_INT_TO_BOOL(OMPI_GROUP_SPARSE);
 bool ompi_use_sparse_group_storage = OPAL_INT_TO_BOOL(OMPI_GROUP_SPARSE);
 
 bool ompi_mpi_yield_when_idle = true;
+bool ompi_mpi_lazy_wait_in_init = false;
 int ompi_mpi_event_tick_rate = -1;
 char *ompi_mpi_show_mca_params_string = NULL;
 bool ompi_mpi_have_sparse_group_storage = !!(OMPI_GROUP_SPARSE);
@@ -112,6 +113,14 @@ int ompi_mpi_register_params(void)
                                  MCA_BASE_VAR_SCOPE_READONLY,
                                  &ompi_mpi_yield_when_idle);
 
+    ompi_mpi_lazy_wait_in_init = false;
+    (void) mca_base_var_register("ompi", "mpi", NULL, "lazy_wait_in_init",
+                                 "Avoid aggressive progress in MPI_Init, make sure that PMIx server has timeslots to progress",
+                                 MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0,
+                                 OPAL_INFO_LVL_9,
+                                 MCA_BASE_VAR_SCOPE_READONLY,
+                                 &ompi_mpi_lazy_wait_in_init);
+
     ompi_mpi_event_tick_rate = -1;
     (void) mca_base_var_register("ompi", "mpi", NULL, "event_tick_rate",
                                  "How often to progress TCP communications (0 = never, otherwise specified in microseconds)",