[Xen-devel] [PATCH 1 of 3] libxl: take node distances into account during NUMA placement

[Dario Faggioli <dario.faggioli@xxxxxxxxxx>]

In fact, among placement candidates with the same number of nodes, the
closer the various nodes are to each others, the better the performances
for a domain placed there.

Signed-off-by: Dario Faggioli <dario.faggioli@xxxxxxxxxx>

diff --git a/tools/libxl/libxl_dom.c b/tools/libxl/libxl_dom.c
--- a/tools/libxl/libxl_dom.c
+++ b/tools/libxl/libxl_dom.c
@@ -105,6 +105,9 @@ out:
  *  - the number of vcpus runnable on the candidates is considered, and
  *    candidates with fewer of them are preferred. If two candidate have
  *    the same number of runnable vcpus,
+ *  - the sum of the node distances in the candidates is considered, and
+ *    candidates with smaller total distance are preferred. If total
+ *    distance is the same for the two candidatess,
  *  - the amount of free memory in the candidates is considered, and the
  *    candidate with greater amount of it is preferred.
  *
@@ -114,6 +117,10 @@ out:
  * overloading large (from a memory POV) nodes. That's right the effect
  * that counting the vcpus able to run on the nodes tries to prevent.
  *
+ * The relative distance within the nodes in the candidates is considered
+ * as the closer the nodes, the better for the domain ending up on the
+ * candidate.
+ *
  * Note that this completely ignore the number of nodes each candidate span,
  * as the fact that fewer nodes is better is already accounted for in the
  * algorithm.
@@ -124,6 +131,9 @@ static int numa_cmpf(const libxl__numa_c
     if (c1->nr_vcpus != c2->nr_vcpus)
         return c1->nr_vcpus - c2->nr_vcpus;
 
+    if (c1->dists_sum != c2->dists_sum)
+        return c1->dists_sum - c2->dists_sum;
+
     return c2->free_memkb - c1->free_memkb;
 }
 
diff --git a/tools/libxl/libxl_internal.h b/tools/libxl/libxl_internal.h
--- a/tools/libxl/libxl_internal.h
+++ b/tools/libxl/libxl_internal.h
@@ -2732,6 +2732,7 @@ static inline void libxl__ctx_unlock(lib
 typedef struct {
     int nr_cpus, nr_nodes;
     int nr_vcpus;
+    int dists_sum;
     uint32_t free_memkb;
     libxl_bitmap nodemap;
 } libxl__numa_candidate;
diff --git a/tools/libxl/libxl_numa.c b/tools/libxl/libxl_numa.c
--- a/tools/libxl/libxl_numa.c
+++ b/tools/libxl/libxl_numa.c
@@ -218,6 +218,40 @@ static int nodemap_to_nr_vcpus(libxl__gc
     return nr_vcpus;
 }
 
+/* Sum the relative distances of nodes in the nodemap to help finding
+ * out which candidate is the "tightest" one. */
+static int nodemap_to_dists_sum(libxl_numainfo *ninfo, libxl_bitmap *nodemap)
+{
+    int tot_dist = 0;
+    int i, j, a = 0, b;
+
+    for (i = 0; i < libxl_bitmap_count_set(nodemap); i++) {
+        while (!libxl_bitmap_test(nodemap, a))
+            a++;
+
+        /* As it is usually non-zero, we do take the latency of
+         * of a node to itself into account. */
+        b = a;
+        for (j = 0; j < libxl_bitmap_count_set(nodemap) - i; j++) {
+            while (!libxl_bitmap_test(nodemap, b))
+                b++;
+
+            /*
+             * In most architectures, going from node A to node B costs
+             * exactly as much as going from B to A does. However, let's
+             * not rely on this and consider both contributions, just to
+             * be ready for everything future might reserve for us.
+             */
+            tot_dist += ninfo[a].dists[b];
+            tot_dist += ninfo[b].dists[a];
+            b++;
+        }
+        a++;
+    }
+
+    return tot_dist;
+}
+
 /*
  * This function tries to figure out if the host has a consistent number
  * of cpus along all its NUMA nodes. In fact, if that is the case, we can
@@ -415,6 +449,7 @@ int libxl__get_numa_candidate(libxl__gc 
              */
             libxl__numa_candidate_put_nodemap(gc, &new_cndt, &nodemap);
             new_cndt.nr_vcpus = nodemap_to_nr_vcpus(gc, tinfo, &nodemap);
+            new_cndt.dists_sum = nodemap_to_dists_sum(ninfo, &nodemap);
             new_cndt.free_memkb = nodes_free_memkb;
             new_cndt.nr_nodes = libxl_bitmap_count_set(&nodemap);
             new_cndt.nr_cpus = nodes_cpus;
@@ -430,12 +465,14 @@ int libxl__get_numa_candidate(libxl__gc 
 
                 LOG(DEBUG, "New best NUMA placement candidate found: "
                            "nr_nodes=%d, nr_cpus=%d, nr_vcpus=%d, "
-                           "free_memkb=%"PRIu32"", new_cndt.nr_nodes,
-                           new_cndt.nr_cpus, new_cndt.nr_vcpus,
+                           "dists_sum=%d, free_memkb=%"PRIu32"",
+                           new_cndt.nr_nodes, new_cndt.nr_cpus,
+                           new_cndt.nr_vcpus, new_cndt.dists_sum,
                            new_cndt.free_memkb / 1024);
 
                 libxl__numa_candidate_put_nodemap(gc, cndt_out, &nodemap);
                 cndt_out->nr_vcpus = new_cndt.nr_vcpus;
+                cndt_out->dists_sum = new_cndt.dists_sum;
                 cndt_out->free_memkb = new_cndt.free_memkb;
                 cndt_out->nr_nodes = new_cndt.nr_nodes;
                 cndt_out->nr_cpus = new_cndt.nr_cpus;

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel