-adding Florian's IBF sim code

diff --git a/src/set/ibf_sim.c b/src/set/ibf_sim.c
new file mode 100644 (file)
index 0000000..61c98dc
--- /dev/null
+++ b/src/set/ibf_sim.c
@@ -0,0 +1,141 @@
+/*
+      This file is part of GNUnet
+      Copyright (C) 2013 Christian Grothoff (and other contributing authors)
+
+      GNUnet is free software; you can redistribute it and/or modify
+      it under the terms of the GNU General Public License as published
+      by the Free Software Foundation; either version 3, or (at your
+      option) any later version.
+
+      GNUnet is distributed in the hope that it will be useful, but
+      WITHOUT ANY WARRANTY; without even the implied warranty of
+      MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+      General Public License for more details.
+
+      You should have received a copy of the GNU General Public License
+      along with GNUnet; see the file COPYING.  If not, write to the
+      Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
+      Boston, MA 02110-1301, USA.
+*/
+
+/**
+ * @file set/ibf_sim.c
+ * @brief implementation of simulation for invertible bloom filter
+ * @author Florian Dold
+ *
+ * This code was used for some internal experiments, it is not
+ * build or shipped as part of the GNUnet system.
+ */
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+
+#define MAX_IBF_DECODE 16
+
+/* report average over how many rounds? */
+#define ROUNDS 100000
+
+/* enable one of the three below */
+// simple fix
+#define FIX1 0
+// possibly slightly better fix for large IBF_DECODE values
+#define FIX2 1
+
+// SIGCOMM algorithm
+#define STRATA 0
+
+// print each value?
+#define VERBOSE 0
+// avoid assembly? (ASM is about 50% faster)
+#define SLOW 0
+
+int
+main(int argc, char **argv)
+{
+  unsigned int round;
+  unsigned int buckets[31]; // max is 2^31 as 'random' returns only between 0 and 2^31
+  unsigned int i;
+  int j;
+  unsigned int r;
+  unsigned int ret;
+  unsigned long long total;
+  unsigned int want;
+  double predict;
+
+  srandom (time (NULL));
+  total = 0;
+  want = atoi (argv[1]);
+  for (round=0;round<ROUNDS;round++)
+  {
+    memset (buckets, 0, sizeof (buckets));
+    for (i=0;i<want;i++)
+    {
+      /* FIXME: might want to use 'better' PRNG to avoid
+        PRNG-induced biases */
+      r = random ();
+      if (0 == r)
+       continue;
+#if SLOW
+      for (j=0;(j < 31) && (0 == (r & (1 << j)));j++) ;
+#else
+      /* use assembly / gcc */
+      j = __builtin_ffs (r) - 1;
+#endif
+      buckets[j]++;
+    }
+    ret = 0;
+    predict = 0.0;
+    for (j=31;j >= 0; j--)
+    {
+#if FIX1
+      /* improved algorithm, for 1000 elements with IBF-DECODE 8, I
+        get 990/1000 elements on average over 1 million runs; key
+        idea being to stop short of the 'last' possible IBF as
+        otherwise a "lowball" per-chance would unduely influence the
+        result */
+      if ( (j > 0) &&
+          (buckets[j - 1] > MAX_IBF_DECODE) )
+      {
+       ret *= (1 << (j + 1));
+       break;
+      }
+#endif
+#if FIX2
+      /* another improvement: don't just always cut off the last one,
+        but rather try to predict based on all previous values where
+        that "last" one is; additional prediction can only really
+        work if MAX_IBF_DECODE is sufficiently high */
+      if (  (j > 0) &&
+           ( (buckets[j - 1] > MAX_IBF_DECODE) ||
+             (predict > MAX_IBF_DECODE) ) )
+      {
+       ret *= (1 << (j + 1));
+       break;
+      }
+#endif
+#if STRATA
+      /* original algorithm, for 1000 elements with IBF-DECODE 8,
+        I get 920/1000 elements on average over 1 million runs */
+      if (buckets[j] > MAX_IBF_DECODE)
+       {
+         ret *= (1 << (j+1));
+         break;
+       }
+#endif
+      ret += buckets[j];
+      predict = (buckets[j] + 2.0 * predict) / 2.0;
+    }
+#if VERBOSE
+    fprintf (stderr, "%u ", ret);
+#endif
+    total += ret;
+  }
+  fprintf (stderr, "\n");
+  fprintf (stdout, "average %llu\n", total / ROUNDS);
+  return 0;
+}
+
+/* TODO: should calculate stddev of the results to also be able to
+   say something about the stability of the results, outside of
+   large-scale averages -- gaining 8% precision at the expense of
+   50% additional variance might not be worth it... */