Once again, the Slurm folks have decided to redefine their envars, reversing what they had previously told us to do. So cleanup the Slurm allocation code, and also adjust to a change in srun behavior that now aborts a job if the ntasks-per-node doesn't get specified when ORTE calls it, but the user specified it when getting an allocation. Sigh.

cmr=v1.7.4:reviewer=miked:subject=Update Slurm allocation and launch

This commit was SVN r29849.

orte/mca/plm/slurm/plm_slurm_module.c

@@ -263,6 +263,9 @@ static void launch_daemons(int fd, short args, void *cbdata)
     /* add the srun command */
     opal_argv_append(&argc, &argv, "srun");
 
+    /* start one orted on each node */
+    opal_argv_append(&argc, &argv, "--ntasks-per-node=1");
+
     /* alert us if any orteds die during startup */
     opal_argv_append(&argc, &argv, "--kill-on-bad-exit");
 

orte/mca/ras/slurm/ras_slurm_module.c

@@ -12,6 +12,7 @@
  * Copyright (c) 2011-2012 Los Alamos National Security, LLC.  All rights
  *                         reserved. 
  * Copyright (c) 2013      Cisco Systems, Inc.  All rights reserved.
+ * Copyright (c) 2013      Intel, Inc. All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -268,9 +269,7 @@ static int orte_ras_slurm_allocate(orte_job_t *jdata, opal_list_t *nodes)
     }
     regexp = strdup(slurm_node_str);
     
-    tasks_per_node = getenv("SLURM_JOB_CPUS_PER_NODE");
-    if (NULL == tasks_per_node) {
-        /* try an older variation */
+    /* get the number of process slots we were assigned on each node */
     tasks_per_node = getenv("SLURM_TASKS_PER_NODE");
     if (NULL == tasks_per_node) {
         /* couldn't find any version - abort */
@@ -278,7 +277,6 @@ static int orte_ras_slurm_allocate(orte_job_t *jdata, opal_list_t *nodes)
                        "SLURM_TASKS_PER_NODE");
         return ORTE_ERR_NOT_FOUND;
     }
-    }
     node_tasks = strdup(tasks_per_node);
 
     if(NULL == regexp || NULL == node_tasks) {