opal/ofi: fix round-robin selection logic

This change fixes current round-robin selection logic:
- Only providers of the same type should be considered, i.e. providers that
match the head of the list. This deviates from the documented behavior.
- For unbound process the selection should be based on its local rank, i.e.
rank among processes on the same node. Currently only the first NIC will be
selected.

Signed-off-by: Wenduo Wang <[email protected]>
(cherry picked from commit b061f96)

Author: wenduwan

opal/mca/common/ofi/common_ofi.c

@@ -771,16 +771,46 @@ static int get_nearest_nic(hwloc_topology_t topology, struct fi_info *provider_l
     return ret;
 }
 
-static struct fi_info *select_provider_round_robin(struct fi_info *provider_list, uint32_t rank,
-                                                   size_t num_providers)
+/**
+ * @brief Selects a provider from the list in a round-robin fashion
+ *
+ * This function implements a round-robin algorithm to select a provider from
+ * the provided list based on a rank. Only providers of the same type as the
+ * first provider are eligible for selection.
+ *
+ * @param[in]   provider_list   A list of providers to select from.
+ * @param[out]  rank            A rank metric for the current process, such as
+ *                              the rank on the same node or CPU package.
+ * @return      Pointer to the selected provider
+ */
+static struct fi_info *select_provider_round_robin(struct fi_info *provider_list, uint32_t rank)
 {
-    uint32_t provider_rank = rank % num_providers;
-    struct fi_info *current_provider = provider_list;
+    uint32_t provider_rank = 0, current_rank = 0;
+    size_t num_providers = 0;
+    struct fi_info *current_provider = NULL;
 
-    for (uint32_t i = 0; i < provider_rank; ++i) {
+    for (current_provider = provider_list; NULL != current_provider;) {
+        if (OPAL_SUCCESS == check_provider_attr(provider_list, current_provider)) {
+            ++num_providers;
+        }
         current_provider = current_provider->next;
     }
 
+    current_provider = provider_list;
+    if (2 > num_providers) {
+        goto out;
+    }
+
+    provider_rank = rank % num_providers;
+
+    while (NULL != current_provider) {
+        if (OPAL_SUCCESS == check_provider_attr(provider_list, current_provider)
+            && provider_rank == current_rank++) {
+            break;
+        }
+        current_provider = current_provider->next;
+    }
+out:
     return current_provider;
 }
 
@@ -888,7 +918,7 @@ struct fi_info *opal_common_ofi_select_provider(struct fi_info *provider_list,
 {
     int ret, num_providers = 0;
     struct fi_info *provider = NULL;
-    uint32_t package_rank = 0;
+    uint32_t package_rank = process_info->my_local_rank;
 
     num_providers = count_providers(provider_list);
     if (!process_info->proc_is_bound || 2 > num_providers) {
@@ -914,7 +944,12 @@ struct fi_info *opal_common_ofi_select_provider(struct fi_info *provider_list,
 #endif /* OPAL_OFI_PCI_DATA_AVAILABLE */
 
 round_robin:
-    provider = select_provider_round_robin(provider_list, package_rank, num_providers);
+    if (!process_info->proc_is_bound && 1 < num_providers
+        && opal_output_get_verbosity(opal_common_ofi.output) >= 1) {
+        opal_show_help("help-common-ofi.txt", "unbound_process", true, 1);
+    }
+
+    provider = select_provider_round_robin(provider_list, package_rank);
 out:
 #if OPAL_ENABLE_DEBUG
     opal_output_verbose(1, opal_common_ofi.output, "package rank: %d device: %s", package_rank,
@@ -988,5 +1023,3 @@ OPAL_DECLSPEC int opal_common_ofi_fi_getname(fid_t fid, void **addr, size_t *add
     }
     return ret;
 }
-
-

opal/mca/common/ofi/help-common-ofi.txt

@@ -7,6 +7,12 @@
 #
 # $HEADER$
 #
+[unbound_process]
+Open MPI's OFI driver detected multiple NICs on the system but cannot select an
+optimal device because the current process is not bound. This may negatively
+impact performance. This can be resolved by specifying "--bind-to ..." on
+command line.
+
 [package_rank failed]
 Open MPI's OFI driver detected multiple equidistant NICs from the current process,
 but had insufficient information to ensure MPI processes fairly pick a NIC for use.