Provide a "progress meter" for launch that outputs progress as we are launching, especially on large jobs. Also, provide a timeout mechanism so that we cleanly abort if we don't get a response from the next daemon in a specified time.

This commit was SVN r21359.

orte/mca/plm/base/plm_base_launch_support.c

@@ -646,6 +646,20 @@ int orte_plm_base_daemon_callback(orte_std_cntr_t num_daemons)
  */
 static bool app_launch_failed;
 static struct timeval max_daemon_launch_msg_recvd = {0,0};
+static orte_vpid_t num_daemons_reported=0;
+static opal_event_t *dmn_report_ev=NULL;
+
+/* catch timeout to allow cmds to progress */
+static void timer_cb(int fd, short event, void *cbdata)
+{
+    /* free event */
+    if (NULL != dmn_report_ev) {
+        free(dmn_report_ev);
+        dmn_report_ev = NULL;
+    }
+    /* declare time is up */
+    app_launch_failed = true;
+}
 
 /* since the HNP also reports launch of procs, we need to separate out
  * the processing of the message vs its receipt so that the HNP
@@ -670,6 +684,13 @@ void orte_plm_base_app_report_launch(int fd, short event, void *data)
                          ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
                          ORTE_NAME_PRINT(&mev->sender)));
     
+    /* got a response - cancel the timer */
+    if (NULL != dmn_report_ev) {
+        opal_event_del(dmn_report_ev);
+        free(dmn_report_ev);
+        dmn_report_ev = NULL;
+    }
+    
     /* unpack the jobid being reported */
     cnt = 1;
     if (ORTE_SUCCESS != (rc = opal_dss.unpack(buffer, &jobid, &cnt, ORTE_JOBID))) {
@@ -689,6 +710,8 @@ void orte_plm_base_app_report_launch(int fd, short event, void *data)
         goto CLEANUP;
     }
     
+    num_daemons_reported++;
+
     /* get the job data object */
     if (NULL == (jdata = orte_get_job_data_object(jobid))) {
         ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
@@ -827,6 +850,14 @@ void orte_plm_base_app_report_launch(int fd, short event, void *data)
         ORTE_ERROR_LOG(rc);
     }
     
+    if (orte_report_launch_progress) {
+        if (0 == num_daemons_reported % 100 || num_daemons_reported == orte_process_info.num_procs) {
+            opal_output(orte_clean_output, "Reported: %d(%d) daemons %d(%d) procs",
+                        (int)num_daemons_reported, (int)orte_process_info.num_procs,
+                        (int)jdata->num_launched, (int)jdata->num_procs);
+        }
+    }
+
     OPAL_OUTPUT_VERBOSE((5, orte_plm_globals.output,
                          "%s plm:base:app_report_launch completed processing",
                          ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
@@ -842,6 +873,10 @@ CLEANUP:
         }
     }
 
+    /* restart the timer, if necessary */
+    if (jdata->num_launched < jdata->num_procs && 0 < orte_startup_timeout) {
+        ORTE_DETECT_TIMEOUT(&dmn_report_ev, orte_startup_timeout, 1000, 10000000, timer_cb);
+    }
 }
 
 
@@ -892,6 +927,13 @@ static int orte_plm_base_report_launched(orte_jobid_t job)
         return ORTE_ERR_NOT_FOUND;
     }
     
+    /* setup a timer - if we don't hear back from a daemon in the
+     * defined time, then we know things have failed
+     */
+    if (0 < orte_startup_timeout) {
+        ORTE_DETECT_TIMEOUT(&dmn_report_ev, orte_startup_timeout, 1000, 10000000, timer_cb);
+    }
+
     /* we should get a callback from every daemon that is involved in
      * the launch. Fortunately, the mapper keeps track of this number
      * for us since num_nodes = num_participating_daemons

orte/runtime/orte_globals.c

@@ -127,6 +127,9 @@ bool orte_assume_same_shell = true;
 /* orted exit with barrier */
 bool orte_orted_exit_with_barrier = true;
 
+/* report launch progress */
+bool orte_report_launch_progress = false;
+
 #endif /* !ORTE_DISABLE_FULL_RTE */
 
 int orte_debug_output = -1;

orte/runtime/orte_globals.h

@@ -514,6 +514,9 @@ ORTE_DECLSPEC extern bool orte_assume_same_shell;
 /* whether or not to barrier the orteds upon exit */
 ORTE_DECLSPEC extern bool orte_orted_exit_with_barrier;
 
+/* whether or not to report launch progress */
+ORTE_DECLSPEC extern bool orte_report_launch_progress;
+
 #endif /* ORTE_DISABLE_FULL_SUPPORT */
 
 END_C_DECLS

orte/runtime/orte_mca_params.c

@@ -284,6 +284,18 @@ int orte_register_params(void)
     mca_base_param_lookup_int(tmp, &value);
     orte_assume_same_shell = OPAL_INT_TO_BOOL(value);
     
+    /* whether or not to report launch progress */
+    mca_base_param_reg_int_name("orte", "report_launch_progress",
+                                "Output a brief periodic report on launch progress [default: no]",
+                                false, false,
+                                (int) false, &value);
+    orte_report_launch_progress = OPAL_INT_TO_BOOL(value);
+    if (orte_report_launch_progress) {
+        /* ensure the startup timeout is set to something reasonable */
+        if (0 == orte_startup_timeout) {
+            orte_startup_timeout = 2000;  /* default to 2 seconds */
+        }
+    }
 #endif /* ORTE_DISABLE_FULL_SUPPORT */
     
     return ORTE_SUCCESS;