mtl/ofi: NIC selection update

The existing code in compare_cpusets assumed that some non_io ancestor of a
PCI object should intersect with the cpuset of the proc. However, this is
not true. There is a case where the non IO ancestor can be an L3. If there
exists two L3s on the same NUMA and the process is bound to one L3, but
the PCI object is connected to the other L3, then compare_cpusets() will
return false.

A better way to determine the optimal interface is by finding the
distances of the interfaces from the current process. Then find out which
of these interfaces is nearest the process and select it.

Use the PMIx distance generation for this purpose.

Signed-off-by: Amir Shehata <[email protected]>

Author: amirshehataornl

opal/mca/common/ofi/common_ofi.c

@@ -445,63 +445,164 @@ static int check_provider_attr(struct fi_info *provider_info, struct fi_info *pr
     }
 }
 
-#if OPAL_OFI_PCI_DATA_AVAILABLE
-/* Check if a process and a pci device share the same cpuset
- *     @param (IN) pci              struct fi_pci_attr pci device attributes,
- *                                  used to find hwloc object for device.
- *
- *     @param (IN) topology         hwloc_topology_t topology to get the cpusets
- *                                  from
- *
- *     @param (OUT)                 returns true if cpusets match and false if
- *                                  cpusets do not match or an error prevents comparison
- *
- *     Uses a pci device to find an ancestor that contains a cpuset, and
- *     determines if it intersects with the cpuset that the process is bound to.
- *     if the process is not bound, or if a cpuset is unavailable for whatever
- *     reason, returns false. Otherwise, returns the result of
- *     hwloc_cpuset_intersects()
- */
-static bool compare_cpusets(hwloc_topology_t topology, struct fi_pci_attr pci)
+static int calculate_distances(pmix_device_distance_t **distances,
+                               size_t *ndist)
 {
-    bool result = false;
-    int ret;
-    hwloc_bitmap_t proc_cpuset;
-    hwloc_obj_t obj = NULL;
+    int ret = 0;
+    size_t ninfo;
+    pmix_info_t *info;
+    pmix_cpuset_t cpuset;
+    pmix_topology_t *pmix_topo;
+    pmix_device_type_t type = PMIX_DEVTYPE_OPENFABRICS |
+      PMIX_DEVTYPE_NETWORK;
+
+    PMIX_CPUSET_CONSTRUCT(&cpuset);
+    ret = PMIx_Get_cpuset(&cpuset, PMIX_CPUBIND_THREAD);
+    if (PMIX_SUCCESS != ret) {
+        goto out;
+    }
+
+    /* load the PMIX topology */
+    PMIX_TOPOLOGY_CREATE(pmix_topo, 1);
+    ret = PMIx_Load_topology(pmix_topo);
+    if (PMIX_SUCCESS != ret) {
+        goto out;
+    }
+
+    ninfo = 1;
+    PMIX_INFO_CREATE(info, ninfo);
+    PMIX_INFO_LOAD(&info[0], PMIX_DEVICE_TYPE, &type, PMIX_DEVTYPE);
+    ret = PMIx_Compute_distances(pmix_topo, &cpuset, info, ninfo, distances,
+                                 ndist);
+    PMIX_INFO_FREE(info, ninfo);
+
+    PMIX_TOPOLOGY_FREE(pmix_topo, 1);
+out:
+    return ret;
+}
 
-    /* Cannot find topology info if no topology is found */
-    if (NULL == topology) {
-        return false;
+static pmix_device_distance_t *get_nearest_nics(pmix_value_t *val,
+                                                int *num_distances)
+{
+    size_t ndist;
+    pmix_topology_t *topo;
+    int ret, i, idx = 0;
+    pmix_data_array_t *dptr;
+    uint16_t near = USHRT_MAX;
+    pmix_info_t directive;
+    pmix_device_distance_t *distances, *nearest = NULL;
+
+    PMIX_INFO_LOAD(&directive, PMIX_OPTIONAL, NULL, PMIX_BOOL);
+    ret = PMIx_Get(&opal_process_info.myprocid,
+             PMIX_DEVICE_DISTANCES, &directive, 1, &val);
+    PMIX_INFO_DESTRUCT(&directive);
+    if (ret != PMIX_SUCCESS || !val) {
+        ret = calculate_distances(&distances, &ndist);
+        if (ret)
+            goto out;
+        goto find_nearest;
+    }
+
+    if (PMIX_DATA_ARRAY != val->type) {
+        goto out;
+    }
+    dptr = val->data.darray;
+    if (NULL == dptr) {
+        goto out;
+    }
+    if (PMIX_DEVICE_DIST != dptr->type) {
+        goto out;
+    }
+
+    distances = (pmix_device_distance_t*)dptr->array;
+    ndist = dptr->size;
+
+find_nearest:
+    nearest = calloc(sizeof(*distances), ndist);
+    if (!nearest)
+        goto out;
+
+    for (i = 0; i < ndist; i++) {
+        if (distances[i].mindist < near) {
+            idx = 0;
+            near = distances[i].mindist;
+            nearest[idx] = distances[i];
+            idx++;
+        } else if (distances[i].mindist == near) {
+            nearest[idx] = distances[i];
+            idx++;
+        }
     }
 
-    /* Allocate memory for proc_cpuset */
-    proc_cpuset = hwloc_bitmap_alloc();
-    if (NULL == proc_cpuset) {
+    *num_distances = idx;
+
+    goto out;
+
+out:
+    return nearest;
+}
+
+#if OPAL_OFI_PCI_DATA_AVAILABLE
+static bool is_near(pmix_device_distance_t *distances,
+                    int num_distances,
+                    hwloc_topology_t topology,
+                    struct fi_pci_attr pci)
+{
+    hwloc_obj_t pcidev, osdev;
+
+    /* if we failed to find any distances, then we consider all interfaces
+     * to be of equal distances and let the caller decide how to handle
+     * them
+     */
+    if (!distances)
+        return true;
+
+    pcidev = hwloc_get_pcidev_by_busid(topology, pci.domain_id,
+                        pci.bus_id, pci.device_id,
+                        pci.function_id);
+    if (!pcidev)
         return false;
-    }
 
-    /* Fill cpuset with the collection of cpu cores that the process runs on */
-    ret = hwloc_get_cpubind(topology, proc_cpuset, HWLOC_CPUBIND_PROCESS);
-    if (0 > ret) {
-        goto error;
-    }
+    for(osdev = pcidev->io_first_child; osdev != NULL; osdev = osdev->next_sibling) {
+        int i;
 
-    /* Get the pci device from bdf */
-    obj = hwloc_get_pcidev_by_busid(topology, pci.domain_id, pci.bus_id, pci.device_id,
-                                    pci.function_id);
-    if (NULL == obj) {
-        goto error;
-    }
+        if (osdev->attr->osdev.type == HWLOC_OBJ_OSDEV_OPENFABRICS) {
+            const char *nguid = hwloc_obj_get_info_by_name(osdev,"NodeGUID");
+            const char *sguid = hwloc_obj_get_info_by_name(osdev, "SysImageGUID");
+
+            if (!nguid && !sguid)
+                continue;
+
+            for (i = 0; i < num_distances; i++) {
+                char lsguid[256], lnguid[256];
+                int ret;
 
-    /* pcidev objects don't have cpusets so find the first non-io object above */
-    obj = hwloc_get_non_io_ancestor_obj(topology, obj);
-    if (NULL != obj) {
-        result = hwloc_bitmap_intersects(proc_cpuset, obj->cpuset);
+                ret = scanf(distances[i].uuid, "fab://%256s::%256s", lnguid, lsguid);
+                if (ret != 2)
+                    continue;
+                if (0 == strcasecmp(lnguid, nguid)) {
+                    return true;
+                } else if (0 == strcasecmp(lsguid, sguid)) {
+                    return true;
+                }
+            }
+        } else if (osdev->attr->osdev.type == HWLOC_OBJ_OSDEV_NETWORK) {
+            const char *address = hwloc_obj_get_info_by_name(osdev, "Address");
+            if (!address)
+                continue;
+            for (i = 0; i < num_distances; i++) {
+                char *addr = strstr(distances[i].uuid, "://");
+                if (!addr || addr + 3 > distances[i].uuid
+                    + strlen(distances[i].uuid))
+                    continue;
+                if (!strcmp(addr+3, address)) {
+                    return true;
+                }
+            }
+        }
     }
 
-error:
-    hwloc_bitmap_free(proc_cpuset);
-    return result;
+    return false;
 }
 #endif
 
@@ -614,7 +715,11 @@ struct fi_info *opal_common_ofi_select_provider(struct fi_info *provider_list,
     struct fi_info *provider = provider_list, *current_provider = provider_list;
     struct fi_info **provider_table;
 #if OPAL_OFI_PCI_DATA_AVAILABLE
+    pmix_device_distance_t *distances = NULL;
+    pmix_value_t *pmix_val = NULL;
     struct fi_pci_attr pci;
+    int num_distances = 0;
+    bool near;
 #endif
     int ret;
     unsigned int num_provider = 0, provider_limit = 0;
@@ -639,33 +744,38 @@ struct fi_info *opal_common_ofi_select_provider(struct fi_info *provider_list,
         return provider_list;
     }
 
+#if OPAL_OFI_PCI_DATA_AVAILABLE
+    distances = get_nearest_nics(pmix_val, &num_distances);
+    if (pmix_val) {
+        PMIX_VALUE_RELEASE(pmix_val);
+    }
+#endif
+
     current_provider = provider;
 
     /* Cycle through remaining fi_info objects, looking for alike providers */
     while (NULL != current_provider) {
         if (!check_provider_attr(provider, current_provider)) {
-            cpusets_match = false;
+            near = false;
 #if OPAL_OFI_PCI_DATA_AVAILABLE
             if (NULL != current_provider->nic
                 && NULL != current_provider->nic->bus_attr
                 && current_provider->nic->bus_attr->bus_type == FI_BUS_PCI) {
                 pci = current_provider->nic->bus_attr->attr.pci;
-                cpusets_match = compare_cpusets(opal_hwloc_topology, pci);
+                near = is_near(distances, num_distances,
+                               opal_hwloc_topology, pci);
             }
 #endif
-
-            /* Reset the list if the cpusets match and no other provider was
-             * found on the same cpuset as the process.
-             */
-            if (cpusets_match && !provider_found) {
+            /* We could have multiple near providers */
+            if (near && !provider_found) {
                 provider_found = true;
                 num_provider = 0;
             }
 
             /* Add the provider to the provider list if the cpusets match or if
              * no other provider was found on the same cpuset as the process.
              */
-            if (cpusets_match || !provider_found) {
+            if (near || !provider_found) {
                 provider_table[num_provider] = current_provider;
                 num_provider++;
             }
@@ -687,17 +797,22 @@ struct fi_info *opal_common_ofi_select_provider(struct fi_info *provider_list,
         && NULL != provider->nic->bus_attr
         && provider->nic->bus_attr->bus_type == FI_BUS_PCI) {
         pci = provider->nic->bus_attr->attr.pci;
-        cpusets_match = compare_cpusets(opal_hwloc_topology, pci);
+        near = is_near(distances, num_distances,
+                       opal_hwloc_topology, pci);
     }
 #endif
 
 #if OPAL_ENABLE_DEBUG
     opal_output_verbose(1, opal_common_ofi.output,
-                        "package rank: %d device: %s cpusets match: %s\n", package_rank,
-                        provider->domain_attr->name, cpusets_match ? "true" : "false");
+                        "package rank: %d device: %s near: %s\n", package_rank,
+                        provider->domain_attr->name, near ? "true" : "false");
 #endif
 
     free(provider_table);
+#if OPAL_OFI_PCI_DATA_AVAILABLE
+    if (distances)
+        free(distances);
+#endif
     return provider;
 }