Implement route selection using a smoothed metric.

This is a systematic way of implementing hyseresis in route
selection, and should solve the issues we're seeing with ETX
jitter and feedback.  The default half-time is 4s, which might
be too small.

babeld.c

@@ -120,9 +120,10 @@ main(int argc, char **argv)
 
     parse_address("ff02:0:0:0:0:0:1:6", protocol_group, NULL);
     protocol_port = 6696;
+    change_smoothing_half_life(4);
 
     while(1) {
-        opt = getopt(argc, argv, "m:p:h:H:i:k:A:PsuS:d:g:lwz:t:T:c:C:DL:I:");
+        opt = getopt(argc, argv, "m:p:h:H:i:k:A:PsuS:d:g:lwz:M:t:T:c:C:DL:I:");
         if(opt < 0)
             break;
 
@@ -212,6 +213,13 @@ main(int argc, char **argv)
                     goto usage;
             }
             break;
+        case 'M': {
+            int l = parse_nat(optarg);
+            if(l < 0 || l > 3600)
+                goto usage;
+            change_smoothing_half_life(l);
+            break;
+        }
         case 't':
             export_table = parse_nat(optarg);
             if(export_table < 0 || export_table > 0xFFFF)
@@ -939,10 +947,10 @@ dump_route_callback(struct babel_route *route, void *closure)
             channels[0] = '\0';
     }
 
-    fprintf(out, "%s metric %d refmetric %d id %s seqno %d%s age %d "
+    fprintf(out, "%s metric %d (%d) refmetric %d id %s seqno %d%s age %d "
             "via %s neigh %s%s%s%s\n",
             format_prefix(route->src->prefix, route->src->plen),
-            route_metric(route), route->refmetric,
+            route_metric(route), route_smoothed_metric(route), route->refmetric,
             format_eui64(route->src->id),
             (int)route->seqno,
             channels,

babeld.man

@@ -62,6 +62,11 @@ The value
 specifies by how much the cost of non-interfering routes is multiplied,
 in units of 1/256; the default is 128 (i.e. division by 2).
 .TP
+.BI \-M " half-time"
+Specify the half-time in seconds of the exponential decay used for
+smoothing metrics for performing route selection; the value 0 disables
+smoothing.  The default is 4s.
+.TP
 .BI \-k " priority"
 Specify the priority value used when installing routes into the kernel.
 The default is 0.

route.c

@@ -48,6 +48,9 @@ int diversity_kind = DIVERSITY_NONE;
 int diversity_factor = 256;     /* in units of 1/256 */
 int keep_unfeasible = 0;
 
+static int smoothing_half_life = 0;
+static int two_to_the_one_over_hl = 0; /* 2^(1/hl) * 0x10000 */
+
 /* We maintain a list of "slots", ordered by prefix.  Every slot
    contains a linked list of the routes to this prefix, with the
    installed route, if any, at the head of the list. */
@@ -481,9 +484,18 @@ change_route_metric(struct babel_route *route,
         }
     }
 
+    /* Update route->smoothed_metric using the old metric. */
+    route_smoothed_metric(route);
+
     route->refmetric = refmetric;
     route->cost = cost;
     route->add_metric = add;
+
+    if(smoothing_half_life == 0) {
+        route->smoothed_metric = route_metric(route);
+        route->smoothed_metric_time = now.tv_sec;
+    }
+
     local_notify_route(route, LOCAL_CHANGE);
 }
 
@@ -570,6 +582,65 @@ update_feasible(struct source *src,
             (src->seqno == seqno && refmetric < src->metric));
 }
 
+void
+change_smoothing_half_life(int half_life)
+{
+    if(half_life <= 0) {
+        smoothing_half_life = 0;
+        two_to_the_one_over_hl = 0;
+        return;
+    }
+
+    smoothing_half_life = half_life;
+    switch(smoothing_half_life) {
+    case 1: two_to_the_one_over_hl = 131072; break;
+    case 2: two_to_the_one_over_hl = 92682; break;
+    case 3: two_to_the_one_over_hl = 82570; break;
+    case 4: two_to_the_one_over_hl = 77935;
+    default:
+        /* 1/2^(1/x) is 1 + log(2)/x + O(1/x^2) at infinity. */
+        two_to_the_one_over_hl = 0x10000 + 45426 / half_life;
+    }
+}
+
+/* Update the smoothed metric, return the new value. */
+int
+route_smoothed_metric(struct babel_route *route)
+{
+    int metric = route_metric(route);
+
+    if(smoothing_half_life <= 0 ||
+       /* Protect against clock stepping. */
+       route->smoothed_metric_time >= now.tv_sec ||
+       route->smoothed_metric == metric) {
+        route->smoothed_metric = metric;
+        route->smoothed_metric_time = now.tv_sec;
+    } else {
+        int diff;
+        /* We randomise the computation, to minimise global synchronisation
+           and hence oscillations. */
+        while(route->smoothed_metric_time <= now.tv_sec - smoothing_half_life) {
+            diff = metric - route->smoothed_metric;
+            route->smoothed_metric += roughly(diff) / 2;
+            route->smoothed_metric_time += smoothing_half_life;
+        }
+        while(route->smoothed_metric_time < now.tv_sec) {
+            diff = metric - route->smoothed_metric;
+            route->smoothed_metric +=
+                roughly(diff) / two_to_the_one_over_hl / 0x10000;
+            route->smoothed_metric_time++;
+        }
+
+        diff = metric - route->smoothed_metric;
+        if(diff > -4 && diff < 4)
+            route->smoothed_metric = metric;
+    }
+
+    /* change_route_metric relies on this */
+    assert(route->smoothed_metric_time == now.tv_sec);
+    return route->smoothed_metric;
+}
+
 static int
 route_acceptable(struct babel_route *route, int feasible,
                  struct neighbour *exclude)
@@ -603,7 +674,7 @@ find_best_route(const unsigned char *prefix, unsigned char plen, int feasible,
     r = route->next;
     while(r) {
         if(route_acceptable(r, feasible, exclude) &&
-           (route_metric(r) < route_metric(route)))
+           (route_smoothed_metric(r) < route_smoothed_metric(route)))
             route = r;
         r = r->next;
     }
@@ -615,6 +686,7 @@ void
 update_route_metric(struct babel_route *route)
 {
     int oldmetric = route_metric(route);
+    int old_smoothed_metric = route_smoothed_metric(route);
 
     if(route_expired(route)) {
         if(route->refmetric < INFINITY) {
@@ -631,7 +703,8 @@ update_route_metric(struct babel_route *route)
                                       neigh->ifp->ifindex);
         change_route_metric(route, route->refmetric,
                             neighbour_cost(route->neigh), add_metric);
-        if(route_metric(route) != oldmetric)
+        if(route_metric(route) != oldmetric ||
+           route_smoothed_metric(route) != old_smoothed_metric)
             route_changed(route, route->src, oldmetric);
     }
 }
@@ -784,6 +857,8 @@ update_route(const unsigned char *id,
         memcpy(route->nexthop, nexthop, 16);
         route->time = now.tv_sec;
         route->hold_time = hold_time;
+        route->smoothed_metric = MAX(route_metric(route), INFINITY / 2);
+        route->smoothed_metric_time = now.tv_sec;
         route->installed = 0;
         memset(&route->channels, 0, sizeof(route->channels));
         if(channels_len > 0)
@@ -854,7 +929,8 @@ consider_route(struct babel_route *route)
     if(route_metric(installed) >= INFINITY)
         goto install;
 
-    if(route_metric(installed) >= route_metric(route) + 64)
+    if(route_metric(installed) >= route_metric(route) &&
+       route_smoothed_metric(installed) > route_smoothed_metric(route))
         goto install;
 
     return;
@@ -960,9 +1036,10 @@ route_changed(struct babel_route *route,
             find_best_route(route->src->prefix, route->src->plen, 1, NULL);
         if(better_route && route_metric(better_route) < route_metric(route))
             consider_route(better_route);
+    }
 
-        if(route->installed)
-            /* We didn't switch to the better route. */
+    if(route->installed) {
+        /* We didn't change routes after all. */
         send_triggered_update(route, oldsrc, oldmetric);
     } else {
         /* Reconsider routes even when their metric didn't decrease,

route.h

@@ -37,6 +37,8 @@ struct babel_route {
     unsigned char nexthop[16];
     time_t time;
     unsigned short hold_time;    /* in seconds */
+    unsigned short smoothed_metric; /* for route selection */
+    time_t smoothed_metric_time;
     short installed;
     unsigned char channels[DIVERSITY_HOPS];
     struct babel_route *next;
@@ -85,6 +87,8 @@ int route_expired(struct babel_route *route);
 int route_interferes(struct babel_route *route, struct interface *ifp);
 int update_feasible(struct source *src,
                     unsigned short seqno, unsigned short refmetric);
+void change_smoothing_half_life(int half_life);
+int route_smoothed_metric(struct babel_route *route);
 struct babel_route *find_best_route(const unsigned char *prefix, unsigned char plen,
                               int feasible, struct neighbour *exclude);
 struct babel_route *install_best_route(const unsigned char prefix[16],