Somehow this has been sitting, uncommitted, in a local checkout since

last December.  :-(

Add new MCA param: maffinity_libnuma_policy.  Thanks to David
Singleton for the suggestion.  Here's the help text about it:

{{{
   MCA maffinity: parameter "maffinity_libnuma_policy" (current value:
                  <loose>, data source: default value)
                  Binding policy that determines what happens if memory
                  is unavailable on the local NUMA node.  A value of
                  "strict" means that the memory allocation will fail;
                  a value of "loose" means that the memory allocation
                  will spill over to another NUMA node.
}}}

This commit was SVN r24290.

opal/mca/maffinity/libnuma/Makefile.am

@@ -9,7 +9,7 @@
 #                         University of Stuttgart.  All rights reserved.
 # Copyright (c) 2004-2005 The Regents of the University of California.
 #                         All rights reserved.
-# Copyright (c) 2010      Cisco Systems, Inc.  All rights reserved.
+# Copyright (c) 2010-2011 Cisco Systems, Inc.  All rights reserved.
 # $COPYRIGHT$
 # 
 # Additional copyrights may follow
@@ -19,6 +19,8 @@
 
 AM_CPPFLAGS = $(maffinity_libnuma_CPPFLAGS)
 
+dist_pkgdata_DATA = help-opal-maffinity-libnuma.txt
+
 sources = \
         maffinity_libnuma.h \
         maffinity_libnuma_component.c \

opal/mca/maffinity/libnuma/help-opal-maffinity-libnuma.txt

@@ -0,0 +1,27 @@
+# -*- text -*-
+#
+# Copyright (c) 2011 Cisco Systems, Inc.  All rights reserved.
+# $COPYRIGHT$
+# 
+# Additional copyrights may follow
+# 
+# $HEADER$
+#
+# This is the US/English help file for Open MPI's libnuma support
+#
+[invalid policy]
+WARNING: An invalid value was given for the maffinity_libnuma_policy
+MCA parameter.  The policy determines what happens when Open MPI tries
+to allocate local memory, but no local memory is available.  The value
+provided was:
+
+  Value:      %s
+  PID:        %d
+
+Valid values are:
+
+  strict: the memory allocation will fail
+  loose:  the memory allocation will spill over to remote memory
+
+Your job will now abort.
+#

opal/mca/maffinity/libnuma/maffinity_libnuma.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) 2011 Cisco Systems, Inc.  All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -34,10 +35,20 @@
 
 BEGIN_C_DECLS
 
+/* Component struct to wrap the base maffinity component and then include
+   some data private to this component. */
+typedef struct {
+    /* Base component */
+    opal_maffinity_base_component_2_0_0_t base;
+
+    /* What libnuma memory binding policy we're using */
+    int libnuma_policy;
+} opal_maffinity_libnuma_component_2_0_0_t;
+
     /**
      * Globally exported variable
      */
-    OPAL_DECLSPEC extern const opal_maffinity_base_component_2_0_0_t
+    OPAL_DECLSPEC extern opal_maffinity_libnuma_component_2_0_0_t
         mca_maffinity_libnuma_component;
 
 

opal/mca/maffinity/libnuma/maffinity_libnuma_component.c

@@ -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) 2011      Cisco Systems, Inc.  All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -18,8 +19,14 @@
 
 #include "opal_config.h"
 
+#include <numa.h>
+#include <numaif.h>
+#include <unistd.h>
+
 #include "opal/constants.h"
+#include "opal/util/show_help.h"
 #include "opal/mca/maffinity/maffinity.h"
+
 #include "maffinity_libnuma.h"
 
 /*
@@ -29,17 +36,22 @@ const char *opal_maffinity_libnuma_component_version_string =
     "OPAL libnuma maffinity MCA component version " OPAL_VERSION;
 
 /*
- * Local function
+ * Local functions
  */
-static int libnuma_open(void);
+static int libnuma_register(void);
+
+/*
+ * Local variable
+ */
+static char *mca_policy = NULL;
 
 /*
  * Instantiate the public struct with all of our public information
  * and pointers to our public functions in it
  */
 
-const opal_maffinity_base_component_2_0_0_t mca_maffinity_libnuma_component = {
+opal_maffinity_libnuma_component_2_0_0_t mca_maffinity_libnuma_component = {
-
+    {
         /* First, the mca_component_t struct containing meta information
            about the component itself */
         
@@ -53,23 +65,48 @@ const opal_maffinity_base_component_2_0_0_t mca_maffinity_libnuma_component = {
             OPAL_RELEASE_VERSION,
             
             /* Component open and close functions */
-        libnuma_open,
             NULL,
-        opal_maffinity_libnuma_component_query
+            NULL,
+            opal_maffinity_libnuma_component_query,
+            libnuma_register,
         },
         {
             /* The component is checkpoint ready */
             MCA_BASE_METADATA_PARAM_CHECKPOINT
         }
+    },
+
+    /* Default libnuma memory binding policy */
+    MPOL_PREFERRED,
 };
 
 
-static int libnuma_open(void)
+static int libnuma_register(void)
 {
-    mca_base_param_reg_int(&mca_maffinity_libnuma_component.base_version,
+    char *val;
+
+    mca_base_param_reg_int(&mca_maffinity_libnuma_component.base.base_version,
                            "priority",
                            "Priority of the libnuma maffinity component",
                            false, false, 25, NULL);
 
+    val = (MPOL_PREFERRED == mca_maffinity_libnuma_component.libnuma_policy ?
+           "loose" : "strict");
+    mca_base_param_reg_string(&mca_maffinity_libnuma_component.base.base_version,
+                              "policy", 
+                              "Binding policy that determines what happens if memory is unavailable on the local NUMA node.  A value of \"strict\" means that the memory allocation will fail; a value of \"loose\" means that the memory allocation will spill over to another NUMA node.",
+                              false, false, val, &mca_policy);
+
+    if (strcasecmp(mca_policy, "loose") == 0) {
+        mca_maffinity_libnuma_component.libnuma_policy = MPOL_PREFERRED;
+    } else if (strcasecmp(mca_policy, "strict") == 0) {
+        mca_maffinity_libnuma_component.libnuma_policy = MPOL_BIND;
+    } else {
+        opal_show_help("help-opal-maffinity-libnuma.txt", "invalid policy",
+                       true, mca_policy, getpid());
+        mca_maffinity_libnuma_component.libnuma_policy = MPOL_PREFERRED;
+        return OPAL_ERR_BAD_PARAM;
+    }
+
     return OPAL_SUCCESS;
 }

opal/mca/maffinity/libnuma/maffinity_libnuma_module.c

@@ -9,7 +9,7 @@
  *                         University of Stuttgart.  All rights reserved.
  * Copyright (c) 2004-2005 The Regents of the University of California.
  *                         All rights reserved.
- * Copyright (c) 2008      Cisco Systems, Inc.  All rights reserved.
+ * Copyright (c) 2008-2011 Cisco Systems, Inc.  All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -69,13 +69,17 @@ int opal_maffinity_libnuma_component_query(mca_base_module_t **module, int *prio
 
 static int libnuma_module_init(void)
 {
-    /* Tell libnuma that we want all memory affinity to be local (but
+    /* Set the libnuma policy.  This affects all memory allocation,
-       it's not an error if we can't -- prefer running in degraded
+       not just libnuma memory allocation. */
-       mode to not running at all!). */
-
-    numa_set_strict(0);
     numa_set_localalloc();
 
+    /* Set strict or not, depending on the value of the MCA param */
+    if (MPOL_BIND == mca_maffinity_libnuma_component.libnuma_policy) {
+        numa_set_strict(1);
+    } else {
+        numa_set_strict(0);
+    }
+
     return OPAL_SUCCESS;
 }
 
@@ -85,12 +89,8 @@ static int libnuma_module_set(opal_maffinity_base_segment_t *segments,
 {
     size_t i;
 
-    /* Kinda crummy that we have to allocate each portion individually
+    /* Explicitly set the memory binding policy for a set of
-       rather than provide a top-level function call that does it all,
+       segments */
-       but the libnuma() interface doesn't seem to allow that
-       flexability -- they allow "interleaving", but not fine grained
-       placement of pages. */
-
     for (i = 0; i < num_segments; ++i) {
         numa_setlocal_memory(segments[i].mbs_start_addr,
                              segments[i].mbs_len);
@@ -115,7 +115,8 @@ static int libnuma_modules_bind(opal_maffinity_base_segment_t *segs,
     unsigned long node_mask = (1 << node_id);
 
     for(i = 0; i < count; i++) {
-        rc = mbind(segs[i].mbs_start_addr, segs[i].mbs_len, MPOL_PREFERRED,
+        rc = mbind(segs[i].mbs_start_addr, segs[i].mbs_len, 
+                   mca_maffinity_libnuma_component.libnuma_policy,
                    &node_mask, sizeof(node_mask) * 8, 
 #ifdef HAVE_MPOL_MF_MOVE
                    MPOL_MF_MOVE