Merge pull request #3801 from rhc54/topic/hetero

Detect that we have a mix of BE/LE in the system

Author: Ralph Castain

configure.ac

@@ -588,7 +588,7 @@ AC_CACHE_SAVE
 opal_show_title "Header file tests"
 
 AC_CHECK_HEADERS([alloca.h aio.h arpa/inet.h dirent.h \
-    dlfcn.h execinfo.h err.h fcntl.h grp.h libgen.h \
+    dlfcn.h endian.h execinfo.h err.h fcntl.h grp.h libgen.h \
     libutil.h memory.h netdb.h netinet/in.h netinet/tcp.h \
     poll.h pthread.h pty.h pwd.h sched.h \
     strings.h stropts.h linux/ethtool.h linux/sockios.h \

opal/mca/hwloc/base/hwloc_base_util.c

@@ -32,6 +32,9 @@
 #ifdef HAVE_UNISTD_H
 #include <unistd.h>
 #endif
+#ifdef HAVE_ENDIAN_H
+#include <endian.h>
+#endif
 
 #include "opal/runtime/opal.h"
 #include "opal/constants.h"
@@ -2155,7 +2158,7 @@ int opal_hwloc_get_sorted_numa_list(hwloc_topology_t topo, char* device_name, op
 char* opal_hwloc_base_get_topo_signature(hwloc_topology_t topo)
 {
     int nnuma, nsocket, nl3, nl2, nl1, ncore, nhwt;
-    char *sig=NULL, *arch=NULL;
+    char *sig=NULL, *arch = NULL, *endian;
     hwloc_obj_t obj;
     unsigned i;
 
@@ -2175,14 +2178,22 @@ char* opal_hwloc_base_get_topo_signature(hwloc_topology_t topo)
             break;
         }
     }
-
     if (NULL == arch) {
-        asprintf(&sig, "%dN:%dS:%dL3:%dL2:%dL1:%dC:%dH",
-                 nnuma, nsocket, nl3, nl2, nl1, ncore, nhwt);
-    } else {
-        asprintf(&sig, "%dN:%dS:%dL3:%dL2:%dL1:%dC:%dH:%s",
-                 nnuma, nsocket, nl3, nl2, nl1, ncore, nhwt, arch);
+        arch = "unknown";
     }
+
+#ifdef __BYTE_ORDER
+#if __BYTE_ORDER == __LITTLE_ENDIAN
+    endian = "le";
+#else
+    endian = "be";
+#endif
+#else
+    endian = "unknown";
+#endif
+
+    asprintf(&sig, "%dN:%dS:%dL3:%dL2:%dL1:%dC:%dH:%s:%s",
+             nnuma, nsocket, nl3, nl2, nl1, ncore, nhwt, arch, endian);
     return sig;
 }
 

orte/mca/plm/base/help-plm-base.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) 2015      Intel, Inc. All rights reserved.
+# Copyright (c) 2015-2017 Intel, Inc. All rights reserved.
 # $COPYRIGHT$
 #
 # Additional copyrights may follow
@@ -162,3 +162,14 @@ A call was made to launch additional processes, but this process has
 no active out-of-band transports and therefore cannot execute this call.
 Please check to see if you have the "oob" MCA parameter set and ensure
 that it is either unset or at least includes the tcp transport.
+#
+[multi-endian]
+Open MPI does not currently support multi-endian operations. We have
+detected that the following node differs in endianness:
+
+
+  Nodename:      %s
+  Endian:        %s
+  Local endian:  %s
+
+Please correct the situation and try again.

orte/mca/plm/base/plm_base_launch_support.c

@@ -1058,12 +1058,23 @@ void orte_plm_base_daemon_callback(int status, orte_process_name_t* sender,
     orte_daemon_cmd_flag_t cmd;
     int32_t flag;
     opal_value_t *kv;
+    char *myendian;
 
     /* get the daemon job, if necessary */
     if (NULL == jdatorted) {
         jdatorted = orte_get_job_data_object(ORTE_PROC_MY_NAME->jobid);
     }
 
+    /* get my endianness */
+    t = (orte_topology_t*)opal_pointer_array_get_item(orte_node_topologies, 0);
+    if (NULL == t) {
+        /* should never happen */
+        myendian = "unknown";
+    } else {
+        myendian = strrchr(t->sig, ':');
+        ++myendian;
+    }
+
     /* multiple daemons could be in this buffer, so unpack until we exhaust the data */
     idx = 1;
     while (OPAL_SUCCESS == (rc = opal_dss.unpack(buffer, &dname, &idx, ORTE_NAME))) {
@@ -1263,8 +1274,24 @@ void orte_plm_base_daemon_callback(int status, orte_process_name_t* sender,
                 }
                 free(sig);
                 break;
+            } else {
+                /* check if the difference is due to the endianness */
+                ptr = strrchr(sig, ':');
+                ++ptr;
+                if (0 != strcmp(ptr, myendian)) {
+                    /* we don't currently handle multi-endian operations in the
+                     * MPI support */
+                    orte_show_help("help-plm-base", "multi-endian", true,
+                                   nodename, ptr, myendian);
+                    orted_failed_launch = true;
+                    if (NULL != topo) {
+                        hwloc_topology_destroy(topo);
+                    }
+                    goto CLEANUP;
+                }
             }
         }
+
         if (!found) {
             /* nope - save the signature and request the complete topology from that node */
             OPAL_OUTPUT_VERBOSE((5, orte_plm_base_framework.framework_output,