Enable slurm operations on Cray with constraints

Cleanup some errors in the nidmap code that caused us to send unnecessary topologies

Signed-off-by: Ralph Castain <rhc@open-mpi.org>

orte/mca/plm/alps/help-plm-alps.txt

@@ -10,6 +10,7 @@
 #                         University of Stuttgart.  All rights reserved.
 # Copyright (c) 2004-2005 The Regents of the University of California.
 #                         All rights reserved.
+# Copyright (c) 2017      Intel, Inc. All rights reserved.
 # $COPYRIGHT$
 #
 # Additional copyrights may follow
@@ -39,7 +40,3 @@ the map for this application. This can be caused by a lack of
 an allocation, or by an error in the Open MPI code. Please check
 to ensure you have a ALPS allocation. If you do, then please pass
 the error to the Open MPI user's mailing list for assistance.
-#
-[slurm-not-supported]
-mpirun is not a supported launcher on Cray XC using Native SLURM.
-srun must be used to launch jobs on these systems.

orte/mca/plm/alps/plm_alps.h

@@ -9,6 +9,7 @@
  *                         University of Stuttgart.  All rights reserved.
  * Copyright (c) 2004-2005 The Regents of the University of California.
  *                         All rights reserved.
+ * Copyright (c) 2017      Intel, Inc. All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -46,7 +47,6 @@ ORTE_MODULE_DECLSPEC extern orte_plm_alps_component_t
     mca_plm_alps_component;
 ORTE_DECLSPEC extern orte_plm_base_module_t
     orte_plm_alps_module;
-extern bool mca_plm_alps_using_aprun;
 
 END_C_DECLS
 

orte/mca/plm/alps/plm_alps_component.c

@@ -12,6 +12,7 @@
  *                         All rights reserved.
  * Copyright (c) 2015      Los Alamos National Security, LLC. All rights
  *                         reserved.
+ * Copyright (c) 2017      Intel, Inc. All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -43,7 +44,6 @@
  */
 const char *mca_plm_alps_component_version_string =
   "Open MPI alps plm MCA component version " ORTE_VERSION;
-bool mca_plm_alps_using_aprun = {true};
 
 
 /*
@@ -158,7 +158,11 @@ static int orte_plm_alps_component_query(mca_base_module_t **module, int *priori
     }
 
     if((NULL != wlm_detected) && !strcmp(slurm, wlm_detected)) {
-        mca_plm_alps_using_aprun = false;
+        /* we are in a Cray SLURM environment, so we don't want
+         * this plm component */
+        *priority = 0;
+        *module = NULL;
+        return ORTE_ERROR;
     }
 #endif
 

orte/mca/plm/alps/plm_alps_module.c

@@ -121,23 +121,6 @@ static int plm_alps_init(void)
         return rc;
     }
 
-    /*
-     * owing to way the SLURM PLM component works, we can't use
-     * it on Cray XC systems as currently designed.  The problem
-     * is the MPI processes launched on the head node (where the
-     * ORTE_PROC_IS_HNP evalues to true) get launched by a daemon
-     * (mpirun) which is not a child of a slurmd daemon.  This
-     * means that any RDMA credentials obtained via the odls/alps
-     * local launcher are incorrect.
-     *
-     * So for now, we just don't support mpirun launched jobs
-     * on Cray XC systems using Native SLURM.
-     */
-    if (false == mca_plm_alps_using_aprun) {
-        orte_show_help("help-plm-alps.txt", "slurm-not-supported", true);
-        exit(-1);
-    }
-
     if (orte_do_not_launch) {
         /* must map daemons since we won't be launching them */
         orte_plm_globals.daemon_nodes_assigned_at_launch = true;

orte/mca/plm/slurm/help-plm-slurm.txt

@@ -10,7 +10,7 @@
 #                         University of Stuttgart.  All rights reserved.
 # Copyright (c) 2004-2005 The Regents of the University of California.
 #                         All rights reserved.
-# Copyright (c) 2014      Intel, Inc. All rights reserved.
+# Copyright (c) 2014-2017 Intel, Inc. All rights reserved.
 # $COPYRIGHT$
 #
 # Additional copyrights may follow
@@ -49,3 +49,18 @@ are running.
 
 Please consult with your system administrator about obtaining
 such support.
+[no-local-support]
+The SLURM process starter cannot start processes local to
+mpirun when executing under a Cray environment. The problem
+is that mpirun is not itself a child of a slurmd daemon. Thus,
+any processes mpirun itself starts will inherit incorrect
+RDMA credentials.
+
+Your application will be mapped and run (assuming adequate
+resources) on the remaining allocated nodes. If adequate
+resources are not available, you will need to exit and obtain
+a larger allocation.
+
+This situation will be fixed in a future release. Meantime,
+you can turn "off" this warning by setting the plm_slurm_warning
+MCA param to 0.

orte/mca/plm/slurm/plm_slurm.h

@@ -9,6 +9,7 @@
  *                         University of Stuttgart.  All rights reserved.
  * Copyright (c) 2004-2005 The Regents of the University of California.
  *                         All rights reserved.
+ * Copyright (c) 2017      Intel, Inc. All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -29,6 +30,7 @@ BEGIN_C_DECLS
 struct orte_plm_slurm_component_t {
     orte_plm_base_component_t super;
     char *custom_args;
+    bool slurm_warning_msg;
 };
 typedef struct orte_plm_slurm_component_t orte_plm_slurm_component_t;
 

orte/mca/plm/slurm/plm_slurm_component.c

@@ -12,6 +12,7 @@
  *                         All rights reserved.
  * Copyright (c) 2015      Los Alamos National Security, LLC. All rights
  *                         reserved.
+ * Copyright (c) 2017      Intel, Inc. All rights reserved.
  * $COPYRIGHT$
  *
  * Additional copyrights may follow
@@ -28,7 +29,9 @@
 #include "orte_config.h"
 #include "orte/constants.h"
 
+#include "opal/util/opal_environ.h"
 #include "orte/util/name_fns.h"
+#include "orte/util/show_help.h"
 #include "orte/runtime/orte_globals.h"
 
 #include "orte/mca/plm/plm.h"
@@ -99,6 +102,13 @@ static int plm_slurm_register(void)
                                             MCA_BASE_VAR_SCOPE_READONLY,
                                             &mca_plm_slurm_component.custom_args);
 
+    mca_plm_slurm_component.slurm_warning_msg = true;
+    (void) mca_base_component_var_register (comp, "warning", "Turn off warning message",
+                                            MCA_BASE_VAR_TYPE_BOOL, NULL, 0, 0,
+                                            OPAL_INFO_LVL_9,
+                                            MCA_BASE_VAR_SCOPE_READONLY,
+                                            &mca_plm_slurm_component.slurm_warning_msg);
+
     return ORTE_SUCCESS;
 }
 

orte/mca/plm/slurm/plm_slurm_module.c

@@ -65,7 +65,7 @@
 #include "orte/runtime/orte_wait.h"
 #include "orte/runtime/orte_quit.h"
 #include "orte/mca/errmgr/errmgr.h"
-#include "orte/mca/rmaps/rmaps.h"
+#include "orte/mca/rmaps/base/base.h"
 #include "orte/mca/state/state.h"
 
 #include "orte/orted/orted.h"
@@ -193,6 +193,25 @@ static void launch_daemons(int fd, short args, void *cbdata)
                          "%s plm:slurm: LAUNCH DAEMONS CALLED",
                          ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
 
+#if SLURM_CRAY_ENV
+        /* if we are in a Cray-SLURM environment, then we cannot
+         * launch procs local to the HNP. The problem
+         * is the MPI processes launched on the head node (where the
+         * ORTE_PROC_IS_HNP evalues to true) get launched by a daemon
+         * (mpirun) which is not a child of a slurmd daemon.  This
+         * means that any RDMA credentials obtained via the odls/alps
+         * local launcher are incorrect. So warn the user and set
+         * the envar for no_schedule_local if mpirun is not on a
+         * system management node (i.e. is part of the allocation)
+         * and the "no_use_local" flag hasn't been set */
+        if (mca_plm_slurm_component.slurm_warning_msg &&
+            (orte_hnp_is_allocated && !(ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping) & ORTE_MAPPING_NO_USE_LOCAL))) {
+            orte_show_help("help-plm-slurm.txt", "no-local-support", true);
+            ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_NO_USE_LOCAL);
+            mca_plm_slurm_component.slurm_warning_msg = false;  // only do this once
+        }
+#endif
+
     /* if we are launching debugger daemons, then just go
      * do it - no new daemons will be launched
      */

orte/util/nidmap.c

@@ -618,14 +618,25 @@ int orte_util_encode_nodemap(opal_buffer_t *buffer)
 
     /* handle the topologies - as the most common case by far
      * is to have homogeneous topologies, we only send them
-     * if something is different */
+     * if something is different. We know that the HNP is
-    if (orte_hnp_is_allocated && !(ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping) & ORTE_MAPPING_NO_USE_LOCAL)) {
+     * the first topology, and that any differing topology
-        ui8 = 2;
+     * on the compute nodes must follow. So send the topologies
-    } else {
+     * if and only if:
-        ui8 = 1;
+     *
+     * (a) the HNP is being used to house application procs and
+     *     there is more than one topology on our list; or
+     *
+     * (b) the HNP is not being used, but there are more than
+     *     two topologies on our list, thus indicating that
+     *     there are multiple topologies on the compute nodes
+     */
+    if (!orte_hnp_is_allocated || (ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping) & ORTE_MAPPING_NO_USE_LOCAL)) {
+        /* remove the first topo on the list */
+        item = opal_list_remove_first(&topos);
+        OBJ_RELEASE(item);
     }
     tmp = NULL;
-    if (ui8 < opal_list_get_size(&topos)) {
+    if (1 < opal_list_get_size(&topos)) {
         opal_buffer_t bucket, *bptr;
         OBJ_CONSTRUCT(&bucket, opal_buffer_t);
         while (NULL != (item = opal_list_remove_first(&topos))) {