Per the discussion on the devel list, move the binding of processes to processors from MPI_Init to process start. This involves:

1. replacing mpi_paffinity_alone with opal_paffinity_alone - for back-compatibility, I have aliased mpi_paffinity_alone to the new param name. This caus
es a mild abstraction break in the opal/mca/paffinity framework - per the devel discussion...live with it. :-) I also moved the ompi_xxx global variable
 that tracked maffinity setup so it could be properly closed in MPI_Finalize to the opal/mca/maffinity framework to avoid an abstraction break.

2. Added code to the odls/default module to perform paffinity binding and maffinity init between process fork and exec. This has been tested on IU's odi
n cluster and works for both MPI and non-MPI apps.

3. Revise MPI_Init to detect if affinity has already been set, and to attempt to set it if not already done. I have *not* tested this as I haven't yet f
igured out a way to do so - I couldn't get slurm to perform cpu bindings, even though it supposedly does do so.

This has only been lightly tested and would definitely benefit from a wider range of evaluation...

This commit was SVN r21209.

ompi/runtime/help-mpi-runtime.txt

@@ -38,7 +38,7 @@ WARNING: Cannot set both the MCA parameters mpi_leave_pinned and
 mpi_leave_pinned_pipeline to "true".  Defaulting to mpi_leave_pinned
 ONLY.
 [mpi_init:startup:paffinity-unavailable]
-The MCA parameter "mpi_paffinity_alone" was set to a nonzero value,
+The MCA parameter "opal_paffinity_alone" was set to a nonzero value,
 but Open MPI was unable to bind MPI_COMM_WORLD rank %s to a processor.
 
 Typical causes for this problem include:

ompi/runtime/mpiruntime.h

@@ -57,10 +57,6 @@ OMPI_DECLSPEC extern int ompi_mpi_thread_provided;
 /** Identifier of the main thread */
 OMPI_DECLSPEC extern struct opal_thread_t *ompi_mpi_main_thread;
 
-/** Did we setup maffinity in MPI_INIT (and therefore need to shut
-    it down during MPI_FINALIZE)? */
-OMPI_DECLSPEC extern bool ompi_mpi_maffinity_setup;
-
 /** Do we want to be warned on fork or not? */
 OMPI_DECLSPEC extern bool ompi_warn_on_fork;
 

ompi/runtime/ompi_mpi_finalize.c

@@ -144,7 +144,7 @@ int ompi_mpi_finalize(void)
     opal_progress_event_users_increment();
 
     /* If maffinity was setup, tear it down */
-    if (ompi_mpi_maffinity_setup) {
+    if (opal_maffinity_setup) {
         opal_maffinity_base_close();
     }
 

ompi/runtime/ompi_mpi_init.c

@@ -280,6 +280,7 @@ int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
     int param, value;
     struct timeval ompistart, ompistop;
     char *event_val = NULL;
+    opal_paffinity_base_cpu_set_t mask;
 #if 0
     /* see comment below about sched_yield */
     int num_processors;
@@ -395,56 +396,60 @@ int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
                           orte_process_info.pid);
     }
 
-    /* Setup process affinity.  First check to see if a slot list was
-       specified.  If so, use it.  If no slot list was specified,
-       that's not an error -- just fall through and try the next
-       paffinity scheme. */
-    ret = opal_paffinity_base_slot_list_set((long)ORTE_PROC_MY_NAME->vpid);
-    if (OPAL_SUCCESS == ret) {
-        paffinity_enabled = true;
-    } 
-    /* If an error occurred in the slot list setup (other than "there
-       was not slot list specified"), bail. */
-    else if (OPAL_ERR_NOT_FOUND != ret) {
-        error = "opal_paffinity_base_slot_list_set() returned an error";
-        goto error;
-    }
-    /* It's an error if multiple paffinity schemes were specified */
-    if (paffinity_enabled && ompi_mpi_paffinity_alone) {
-        ret = OMPI_ERR_BAD_PARAM;
-        error = "Multiple processor affinity schemes specified (can only specify one)";
-        goto error;
-    } 
-    /* Otherwise, if mpi_paffinity_alone was set, use that scheme */
-    else if (ompi_mpi_paffinity_alone) {
-        opal_paffinity_base_cpu_set_t mask;
-        int phys_cpu;
-        orte_node_rank_t nrank;
-        if (ORTE_NODE_RANK_INVALID == (nrank = orte_ess.get_node_rank(ORTE_PROC_MY_NAME))) {
-            error = "Could not get node rank - cannot set processor affinity";
-            goto error;
+    /* if it hasn't already been done, setup process affinity. 
+     * First check to see if a slot list was
+     * specified.  If so, use it.  If no slot list was specified,
+     * that's not an error -- just fall through and try the next
+     * paffinity scheme.
+     */
+    ret = opal_paffinity_base_get(&mask);
+    if (OPAL_ERR_NOT_FOUND == ret) {
+        /* the system is capable of doing processor affinity, but it
+         * has not yet been set - see if a slot_list was given
+         */
+        if (NULL != opal_paffinity_base_slot_list) {
+            /* It's an error if multiple paffinity schemes were specified */
+            if (opal_paffinity_alone) {
+                ret = OMPI_ERR_BAD_PARAM;
+                error = "Multiple processor affinity schemes specified (can only specify one)";
+                goto error;
+            }
+            ret = opal_paffinity_base_slot_list_set((long)ORTE_PROC_MY_NAME->vpid, opal_paffinity_base_slot_list);
+            if (OPAL_ERR_NOT_FOUND != ret) {
+                error = "opal_paffinity_base_slot_list_set() returned an error";
+                goto error;
+            }
+            paffinity_enabled = true;
+        } else if (opal_paffinity_alone) {
+            /* no slot_list, but they asked for paffinity */
+            int phys_cpu;
+            orte_node_rank_t nrank;
+            if (ORTE_NODE_RANK_INVALID == (nrank = orte_ess.get_node_rank(ORTE_PROC_MY_NAME))) {
+                error = "Could not get node rank - cannot set processor affinity";
+                goto error;
+            }
+            OPAL_PAFFINITY_CPU_ZERO(mask);
+            phys_cpu = opal_paffinity_base_get_physical_processor_id(nrank);
+            if (0 > phys_cpu) {
+                error = "Could not get physical processor id - cannot set processor affinity";
+                goto error;
+            }
+            OPAL_PAFFINITY_CPU_SET(phys_cpu, mask);
+            ret = opal_paffinity_base_set(mask);
+            if (OPAL_SUCCESS != ret) {
+                error = "Setting processor affinity failed";
+                goto error;
+            }
+            paffinity_enabled = true;
         }
-        OPAL_PAFFINITY_CPU_ZERO(mask);
-        phys_cpu = opal_paffinity_base_get_physical_processor_id(nrank);
-        if (0 > phys_cpu) {
-            error = "Could not get physical processor id - cannot set processor affinity";
-            goto error;
-        }
-        OPAL_PAFFINITY_CPU_SET(phys_cpu, mask);
-        ret = opal_paffinity_base_set(mask);
-        if (OPAL_SUCCESS != ret) {
-            error = "Setting processor affinity failed";
-            goto error;
-        }
-        paffinity_enabled = true;
     }
 
     /* If we were able to set processor affinity, try setting up
        memory affinity */
-    if (paffinity_enabled) {
+    if (!opal_maffinity_setup && paffinity_enabled) {
         if (OPAL_SUCCESS == opal_maffinity_base_open() &&
             OPAL_SUCCESS == opal_maffinity_base_select()) {
-            ompi_mpi_maffinity_setup = true;
+            opal_maffinity_setup = true;
         }
     }
     

ompi/runtime/ompi_mpi_params.c

@@ -52,7 +52,6 @@ int ompi_debug_show_mpi_alloc_mem_leaks = 0;
 bool ompi_debug_no_free_handles = false;
 bool ompi_mpi_show_mca_params = false;
 char *ompi_mpi_show_mca_params_file = NULL;
-bool ompi_mpi_paffinity_alone = false;
 bool ompi_mpi_abort_print_stack = false;
 int ompi_mpi_abort_delay = 0;
 bool ompi_mpi_keep_peer_hostnames = true;
@@ -261,12 +260,6 @@ int ompi_mpi_register_params(void)
                        true);
     }
 
-    mca_base_param_reg_int_name("mpi", "paffinity_alone",
-                                "If nonzero, assume that this job is the only (set of) process(es) running on each node and bind processes to processors, starting with processor ID 0",
-                                false, false, 
-                                (int) ompi_mpi_paffinity_alone, &value);
-    ompi_mpi_paffinity_alone = OPAL_INT_TO_BOOL(value);
-
     mca_base_param_reg_int_name("mpi", "warn_on_fork",
                                 "If nonzero, issue a warning if program forks under conditions that could cause system errors",
                                 false, false, 

ompi/runtime/params.h

@@ -95,15 +95,6 @@ OMPI_DECLSPEC extern bool ompi_mpi_show_mca_params;
  */
 OMPI_DECLSPEC extern char * ompi_mpi_show_mca_params_file;
 
-/**
- * If this value is true, assume that this ORTE job is the only job
- * running on the nodes that have been allocated to it, and bind
- * processes to the processor ID corresponding to their node local
- * rank (if you COMM_SPAWN on to empty processors on the same node,
- * the NLR will start at N, not 0).
- */
-OMPI_DECLSPEC extern bool ompi_mpi_paffinity_alone;
-
 /**
  * Whether we should keep the string hostnames of all the MPI
  * process peers around or not (eats up a good bit of memory).

opal/mca/maffinity/base/base.h

@@ -154,7 +154,12 @@ OPAL_DECLSPEC extern opal_list_t opal_maffinity_base_components_opened;
  * Debugging output stream
  */
 extern int opal_maffinity_base_output;
-    
+
+/**
+ * Flag to indicate whether or not maffinity was setup
+ */
+OPAL_DECLSPEC extern bool opal_maffinity_setup;
+
 END_C_DECLS
     
 #endif /* OPAL_BASE_MAFFINITY_H */

opal/mca/maffinity/base/maffinity_base_open.c

@@ -42,7 +42,7 @@
 int opal_maffinity_base_output = -1;
 bool opal_maffinity_base_components_opened_valid = false;
 opal_list_t opal_maffinity_base_components_opened;
-
+bool opal_maffinity_setup = false;
 
 /*
  * Function for finding and opening either all MCA components, or the one

opal/mca/maffinity/maffinity_types.h

@@ -28,36 +28,32 @@
 
 #include <sys/types.h>
 
-#if defined(c_plusplus) || defined(__cplusplus)
-extern "C" {
-#endif
+BEGIN_C_DECLS
 
-    /**
-     * Struct used with opal_maffinity_base_module_set_fn_t.  It
-     * describes a section of memory (starting address and length).
-     * This is really the same thing as an iovec, but we include a
-     * separate type for it for at least 2 reasons:
-     *
-     * 1. Some OS's iovec definitions are exceedingly lame (e.g.,
-     * Solaris 9 has the length argument as an int, instead of a
-     * size_t).
-     *
-     * 2. We reserve the right to expand/change this struct in the
-     * future.
-     */
-    struct opal_maffinity_base_segment_t {
-        /** Starting address of segment */
-        void *mbs_start_addr;
-        /** Length of segment */
-        size_t mbs_len;
-    };
-    /**
-     * Convenience typedef
-     */
-    typedef struct opal_maffinity_base_segment_t opal_maffinity_base_segment_t;
+/**
+ * Struct used with opal_maffinity_base_module_set_fn_t.  It
+ * describes a section of memory (starting address and length).
+ * This is really the same thing as an iovec, but we include a
+ * separate type for it for at least 2 reasons:
+ *
+ * 1. Some OS's iovec definitions are exceedingly lame (e.g.,
+ * Solaris 9 has the length argument as an int, instead of a
+ * size_t).
+ *
+ * 2. We reserve the right to expand/change this struct in the
+ * future.
+ */
+struct opal_maffinity_base_segment_t {
+    /** Starting address of segment */
+    void *mbs_start_addr;
+    /** Length of segment */
+    size_t mbs_len;
+};
+/**
+ * Convenience typedef
+ */
+typedef struct opal_maffinity_base_segment_t opal_maffinity_base_segment_t;
 
-#if defined(c_plusplus) || defined(__cplusplus)
-}
-#endif
+END_C_DECLS
 
 #endif /* OPAL_MAFFINITY_TYPES_H */

opal/mca/paffinity/base/base.h

@@ -237,13 +237,20 @@ OPAL_DECLSPEC extern opal_list_t opal_paffinity_base_components_opened;
 /**
  * Assigning slot_list to process 
  */
-OPAL_DECLSPEC int opal_paffinity_base_slot_list_set(long rank);
+OPAL_DECLSPEC int opal_paffinity_base_slot_list_set(long rank, char *slot_str);
 
 /**
  * Debugging output stream
  */
 OPAL_DECLSPEC extern int opal_paffinity_base_output;
 
+/**
+ * Flag indicating whether or not processor affinity is to be enabled
+ */
+OPAL_DECLSPEC extern bool opal_paffinity_alone;
+
+OPAL_DECLSPEC extern char *opal_paffinity_base_slot_list;
+
 END_C_DECLS
 
 #endif /* OPAL_BASE_PAFFINITY_H */

opal/mca/paffinity/base/paffinity_base_open.c

@@ -43,7 +43,8 @@
 OPAL_DECLSPEC int opal_paffinity_base_output = -1;
 bool opal_paffinity_base_components_opened_valid = false;
 opal_list_t opal_paffinity_base_components_opened;
-
+bool opal_paffinity_alone = false;
+char *opal_paffinity_base_slot_list;
 
 /*
  * Function for finding and opening either all MCA components, or the one
@@ -51,7 +52,7 @@ opal_list_t opal_paffinity_base_components_opened;
  */
 int opal_paffinity_base_open(void)
 {
-    int value;
+    int value, id;
 
     /* Debugging / verbose output */
 
@@ -65,11 +66,19 @@ int opal_paffinity_base_open(void)
         opal_paffinity_base_output = -1;
     }
 
+    id = mca_base_param_reg_int_name("opal", "paffinity_alone", 
+                                "If nonzero, assume that this job is the only (set of) process(es) running on each node and bind processes to processors, starting with processor ID 0",
+                                false, false,
+                                0, NULL);
+    mca_base_param_reg_syn_name(id, "mpi", "paffinity_alone", true);
+    mca_base_param_lookup_int(id, &value);
+    opal_paffinity_alone = OPAL_INT_TO_BOOL(value);
+
     opal_paffinity_base_components_opened_valid = false;
         
     mca_base_param_reg_string_name("opal", "paffinity_base_slot_list",
                                    "Used to set list of processor IDs to bind MPI processes to (e.g., used in conjunction with rank files)",
-                                   true, false, NULL, NULL);
+                                   true, false, NULL, &opal_paffinity_base_slot_list);
 
     /* Open up all available components */
 

opal/mca/paffinity/base/paffinity_base_service.c

@@ -512,29 +512,22 @@ static int opal_paffinity_base_socket_core_to_cpu_set(char **socket_core_list, i
     return OPAL_SUCCESS;
 }
 
-int opal_paffinity_base_slot_list_set(long rank)
+int opal_paffinity_base_slot_list_set(long rank, char *slot_str)
 {
-    char *slot_str = NULL;
     char **item;
     char **socket_core;
     int item_cnt, socket_core_cnt, rc;
     bool logical_map;
     
-    rc = mca_base_param_find("opal", NULL, "paffinity_base_slot_list");
-    /* If there was not slot list specified, return a specific error
-     code indicating that */
-    if (rc <= 0) {
-        return OPAL_ERR_NOT_FOUND;
-    }
-    
-    if (OPAL_SUCCESS == mca_base_param_lookup_string(rc, &slot_str)) {
-        if (NULL == slot_str) {
-            return OPAL_ERR_NOT_FOUND;
-        }
-    }
-    if (0 == strcmp("", slot_str)){
+    if (NULL == slot_str){
         return OPAL_ERR_BAD_PARAM;
     }
+    
+    /* if the slot string is empty, that is an error */
+    if (0 == strlen(slot_str)) {
+        return OPAL_ERR_BAD_PARAM;
+    }
+    
     /* check for diag request to avoid repeatedly doing so */
     if (4 < opal_output_get_verbosity(opal_paffinity_base_output)) {
         diag_requested = true;

orte/mca/odls/base/odls_base_default_fns.c

@@ -1285,16 +1285,6 @@ int orte_odls_base_default_launch_local(orte_jobid_t job,
             opal_setenv("OMPI_COMM_WORLD_LOCAL_RANK", value, true, &app->env);
             free(value);
             
-           param = mca_base_param_environ_variable("opal", NULL, "paffinity_base_slot_list");
-            if ( NULL != child->slot_list ) {
-                asprintf(&value, "%s", child->slot_list);
-                opal_setenv(param, value, true, &app->env);
-                free(value);
-            } else {
-                opal_unsetenv(param,  &app->env);
-            }
-            free(param);
-            
            /* if we are timing things, record when we are going to launch this proc */
             if (orte_timing) {
                 gettimeofday(&child->starttime, NULL);

orte/mca/odls/default/help-odls-default.txt

@@ -55,3 +55,36 @@ Max value allowed: %ld
 This may be resolved by increasing the number of available node id's by
 re-configuring Open MPI with the --enable-jumbo-clusters option, and then
 re-running the application
+#
+[odls-default:multiple-paffinity-schemes]
+Multiple processor affinity schemes were specified (can only specify one):
+
+Slot list: %s
+opal_paffinity_alone:  true
+
+Please specify only the one desired method.
+#
+[odls-default:slot-list-failed]
+We were unable to successfully process/set the requested processor
+affinity settings:
+
+Specified slot list: %s
+Error: %s
+
+This could mean that a non-existent processor was specified, or
+that the specification had improper syntax.
+#
+[odls-default:invalid-node-rank]
+An invalid node rank was obtained - this is probably something
+that should be reported to the OMPI developers.
+#
+[odls-default:invalid-phys-cpu]
+An invalid physical processor id was returned when attempting to
+set processor affinity. This is probably something that should be
+reported to the OMPI developers - your system may not support
+this functionality.
+#
+[odls-default:failed-set-paff]
+An attempt to set processor affinity has failed - please check to
+ensure that your system supports such functionality. If so, then
+this is probably something that should be reported to the OMPI developers.

orte/mca/odls/default/odls_default_module.c

@@ -67,10 +67,14 @@
 #endif
 #endif /* HAVE_SCHED_YIELD */
 
+#include "opal/mca/maffinity/base/base.h"
+#include "opal/mca/paffinity/base/base.h"
+
 #include "orte/util/show_help.h"
 #include "orte/runtime/orte_wait.h"
 #include "orte/runtime/orte_globals.h"
 #include "orte/mca/errmgr/errmgr.h"
+#include "orte/mca/ess/ess.h"
 #include "orte/mca/iof/base/iof_base_setup.h"
 #include "orte/util/name_fns.h"
 
@@ -178,6 +182,7 @@ static int odls_default_fork_local_proc(orte_app_context_t* context,
     sigset_t sigs;
     int i, p[2];
     pid_t pid;
+    bool paffinity_enabled = false;
 
     if (NULL != child) {
         /* should pull this information from MPIRUN instead of going with
@@ -259,7 +264,75 @@ static int odls_default_fork_local_proc(orte_app_context_t* context,
                 exit(1);
             }
             
-            
+           /* Setup process affinity.  First check to see if a slot list was
+            * specified.  If so, use it.  If no slot list was specified,
+            * that's not an error -- just fall through and try the next
+            * paffinity scheme.
+            */
+           if (NULL != child->slot_list) {
+               OPAL_OUTPUT_VERBOSE((2, orte_odls_globals.output,
+                                    "%s odls:default:fork got slot_list %s for child %s",
+                                    ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
+                                    child->slot_list, ORTE_NAME_PRINT(child->name)));
+               if (opal_paffinity_alone) {
+                   /* It's an error if multiple paffinity schemes were specified */
+                   orte_show_help("help-odls-default.txt",
+                                  "odls-default:multiple-paffinity-schemes", true, child->slot_list);
+                   rc = ORTE_ERR_FATAL;
+                   write(p[1], &rc, sizeof(int));
+                   exit(1);
+               }
+               if (OPAL_SUCCESS != (rc = opal_paffinity_base_slot_list_set((long)child->name->vpid, child->slot_list))) {
+                   orte_show_help("help-odls-default.txt",
+                                  "odls-default:slot-list-failed", true, child->slot_list, ORTE_ERROR_NAME(rc));
+                   write(p[1], &rc, sizeof(int));
+                   exit(1);
+               }
+           }
+           /* Otherwise, if opal_paffinity_alone was set, use that scheme */
+           else if (opal_paffinity_alone) {
+               opal_paffinity_base_cpu_set_t mask;
+               int phys_cpu;
+               orte_node_rank_t nrank;
+               OPAL_OUTPUT_VERBOSE((2, orte_odls_globals.output,
+                                    "%s odls:default:fork setting paffinity for child %s",
+                                    ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
+                                    ORTE_NAME_PRINT(child->name)));
+               if (ORTE_NODE_RANK_INVALID == (nrank = orte_ess.get_node_rank(child->name))) {
+                   orte_show_help("help-odls-default.txt",
+                                  "odls-default:invalid-node-rank", true);
+                   rc = ORTE_ERR_FATAL;
+                   write(p[1], &rc, sizeof(int));
+                   exit(1);
+               }
+               OPAL_PAFFINITY_CPU_ZERO(mask);
+               phys_cpu = opal_paffinity_base_get_physical_processor_id(nrank);
+               if (0 > phys_cpu) {
+                   orte_show_help("help-odls-default.txt",
+                                  "odls-default:invalid-phys-cpu", true);
+                   rc = ORTE_ERR_FATAL;
+                   write(p[1], &rc, sizeof(int));
+                   exit(1);
+               }
+               OPAL_PAFFINITY_CPU_SET(phys_cpu, mask);
+               if (OPAL_SUCCESS != (rc = opal_paffinity_base_set(mask))) {
+                   orte_show_help("help-odls-default.txt",
+                                  "odls-default:failed-set-paff", true);
+                   write(p[1], &rc, sizeof(int));
+                   exit(1);
+               }
+               paffinity_enabled = true;
+           }
+           /* If we were able to set processor affinity, try setting up
+            * memory affinity
+            */
+           if (paffinity_enabled) {
+               if (OPAL_SUCCESS == opal_maffinity_base_open() &&
+                   OPAL_SUCCESS == opal_maffinity_base_select()) {
+                   opal_maffinity_setup = true;
+               }
+           }
+           
         } else if (!(ORTE_JOB_CONTROL_FORWARD_OUTPUT & controls)) {
             /* tie stdin/out/err/internal to /dev/null */
             int fdnull;