Bug 1335: Alter Xm calculation to be weighted avg of top N=3 modes.

In my state files, I was seeing several peaks, probably due to different
guards having different latency. This change is meant to better capture
this behavior and generate more reasonable timeouts when it happens. It
is improving the timeout values for my collection of state files.

1 files changed, 40 insertions, 10 deletions

diff --git a/src/or/circuitbuild.c b/src/or/circuitbuild.c
index 233d60f15..98f6baf2b 100644
--- a/src/or/circuitbuild.c
+++ b/src/or/circuitbuild.c
@@ -20,6 +20,8 @@
 #define MIN(a,b) ((a)<(b)?(a):(b))
 #endif
 
+#define CBT_BIN_TO_MS(bin) ((bin)*CBT_BIN_WIDTH + (CBT_BIN_WIDTH/2))
+
 /********* START VARIABLES **********/
 /** Global list of circuit build times */
 // FIXME: Add this as a member for entry_guard_t instead of global?
@@ -387,25 +389,53 @@ circuit_build_times_create_histogram(circuit_build_times_t *cbt,
 }
 
 /**
- * Return the most frequent build time (rounded to CBT_BIN_WIDTH ms).
+ * Return the Pareto start-of-curve parameter Xm.
  *
- * Ties go in favor of the slower time.
+ * Because we are not a true Pareto curve, we compute this as the
+ * weighted average of the N=3 most frequent build time bins.
  */
 static build_time_t
-circuit_build_times_mode(circuit_build_times_t *cbt)
+circuit_build_times_get_xm(circuit_build_times_t *cbt)
 {
-  build_time_t i, nbins, max_bin=0;
+  build_time_t i, nbins;
+  build_time_t nth_max_bin[CBT_NUM_XM_MODES];
+  int32_t bin_counts=0;
+  build_time_t ret = 0;
   uint32_t *histogram = circuit_build_times_create_histogram(cbt, &nbins);
+  int n=0;
+  int num_modes = CBT_NUM_XM_MODES;
+
+  // Only use one mode if < 1000 buildtimes. Not enough data
+  // for multiple.
+  if (cbt->total_build_times < CBT_NCIRCUITS_TO_OBSERVE)
+    num_modes = 1;
 
+  memset(nth_max_bin, 0, sizeof(nth_max_bin));
   for (i = 0; i < nbins; i++) {
-    if (histogram[i] >= histogram[max_bin]) {
-      max_bin = i;
+    if (histogram[i] >= histogram[nth_max_bin[0]]) {
+      nth_max_bin[0] = i;
     }
+
+    for (n = 1; n < num_modes; n++) {
+      if (histogram[i] >= histogram[nth_max_bin[n]] &&
+           (!histogram[nth_max_bin[n-1]]
+               || histogram[i] < histogram[nth_max_bin[n-1]])) {
+        nth_max_bin[n] = i;
+      }
+    }
+  }
+
+  for (n = 0; n < num_modes; n++) {
+    bin_counts += histogram[nth_max_bin[n]];
+    ret += CBT_BIN_TO_MS(nth_max_bin[n])*histogram[nth_max_bin[n]];
+    log_info(LD_CIRC, "Xm mode #%d: %u %u", n, CBT_BIN_TO_MS(nth_max_bin[n]),
+             histogram[nth_max_bin[n]]);
   }
 
+  ret /= bin_counts;
   tor_free(histogram);
 
-  return max_bin*CBT_BIN_WIDTH+CBT_BIN_WIDTH/2;
+  return ret;
 }
 
 /**
@@ -436,7 +466,7 @@ circuit_build_times_update_state(circuit_build_times_t *cbt,
     line->key = tor_strdup("CircuitBuildTimeBin");
     line->value = tor_malloc(25);
     tor_snprintf(line->value, 25, "%d %d",
-            i*CBT_BIN_WIDTH+CBT_BIN_WIDTH/2, histogram[i]);
+            CBT_BIN_TO_MS(i), histogram[i]);
     next = &(line->next);
   }
 
@@ -596,7 +626,7 @@ circuit_build_times_update_alpha(circuit_build_times_t *cbt)
   /* http://en.wikipedia.org/wiki/Pareto_distribution#Parameter_estimation */
   /* We sort of cheat here and make our samples slightly more pareto-like
    * and less frechet-like. */
-  cbt->Xm = circuit_build_times_mode(cbt);
+  cbt->Xm = circuit_build_times_get_xm(cbt);
 
   for (i=0; i< CBT_NCIRCUITS_TO_OBSERVE; i++) {
     if (!x[i]) {
@@ -763,7 +793,7 @@ circuit_build_times_count_pretimeouts(circuit_build_times_t *cbt)
                     (cbt->pre_timeouts+cbt->total_build_times);
     /* Make sure it doesn't exceed the synthetic max */
     timeout_quantile *= CBT_MAX_SYNTHETIC_QUANTILE;
-    cbt->Xm = circuit_build_times_mode(cbt);
+    cbt->Xm = circuit_build_times_get_xm(cbt);
     tor_assert(cbt->Xm > 0);
     /* Use current timeout to get an estimate on alpha */
     circuit_build_times_initial_alpha(cbt, timeout_quantile,