fixed issue with grpcomm rcd and brks algorithms which led to performance issues: data just for part of processes was unpacked and stored locally during fence, therefore clients were forced to ask daemons for data directly during get request

orte/mca/grpcomm/brks/grpcomm_brks.c

@@ -307,6 +307,8 @@ static void brks_allgather_recv_dist(int status, orte_process_name_t* sender,
 static int brks_finalize_coll(orte_grpcomm_coll_t *coll, int ret) {
     opal_buffer_t *reply;
     int rc;
+    orte_job_t *jdata;
+    uint64_t nprocs;
 
     OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base_framework.framework_output,
                          "%s grpcomm:coll:brks declared collective complete",
@@ -314,11 +316,25 @@ static int brks_finalize_coll(orte_grpcomm_coll_t *coll, int ret) {
 
     reply = OBJ_NEW(opal_buffer_t);
 
-    if (OPAL_SUCCESS != (rc = opal_dss.pack(reply, &coll->nreported, 1, OPAL_UINT64))) {
+    /* pack the number of procs involved in the collective
+     * so the recipients can unpack any collected data */
+    if (1 == coll->sig->sz) {
+        /* get the job object for this entry */
+        if (NULL == (jdata = orte_get_job_data_object(coll->sig->signature[0].jobid))) {
+            ORTE_ERROR_LOG(rc);
+            OBJ_RELEASE(reply);
+            return rc;
+        }
+        nprocs = jdata->num_procs;
+    } else {
+        nprocs = coll->sig->sz;
+    }
+    if (OPAL_SUCCESS != (rc = opal_dss.pack(reply, &nprocs, 1, OPAL_UINT64))) {
         ORTE_ERROR_LOG(rc);
         OBJ_RELEASE(reply);
         return rc;
     }
+
     /* transfer the collected bucket */
     opal_dss.copy_payload(reply, &coll->bucket);
 

orte/mca/grpcomm/rcd/grpcomm_rcd.c

@@ -311,6 +311,8 @@ static void rcd_allgather_recv_dist(int status, orte_process_name_t* sender,
 static int rcd_finalize_coll(orte_grpcomm_coll_t *coll, int ret) {
     opal_buffer_t *reply;
     int rc;
+    orte_job_t *jdata;
+    uint64_t nprocs;
 
     OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base_framework.framework_output,
                          "%s grpcomm:coll:recdub declared collective complete",
@@ -318,11 +320,25 @@ static int rcd_finalize_coll(orte_grpcomm_coll_t *coll, int ret) {
 
     reply = OBJ_NEW(opal_buffer_t);
 
-    if (OPAL_SUCCESS != (rc = opal_dss.pack(reply, &coll->nreported, 1, OPAL_UINT64))) {
+    /* pack the number of procs involved in the collective
+     * so the recipients can unpack any collected data */
+    if (1 == coll->sig->sz) {
+        /* get the job object for this entry */
+        if (NULL == (jdata = orte_get_job_data_object(coll->sig->signature[0].jobid))) {
+            ORTE_ERROR_LOG(rc);
+            OBJ_RELEASE(reply);
+            return rc;
+        }
+        nprocs = jdata->num_procs;
+    } else {
+        nprocs = coll->sig->sz;
+    }
+    if (OPAL_SUCCESS != (rc = opal_dss.pack(reply, &nprocs, 1, OPAL_UINT64))) {
         ORTE_ERROR_LOG(rc);
         OBJ_RELEASE(reply);
         return rc;
     }
+
     /* transfer the collected bucket */
     opal_dss.copy_payload(reply, &coll->bucket);