MINOR: freq_ctr: introduce a new averaging method

While the current functions report average event counts per period, we are
also interested in average values per event. For this we use a different
method. The principle is to rely on a long tail which sums the new value
with a fraction of the previous value, resulting in a sliding window of
infinite length depending on the precision we're interested in.

The idea is that we always keep (N-1)/N of the sum and add the new sampled
value. The sum over N values can be computed with a simple program for a
constant value 1 at each iteration :

    N
  ,---
   \       N - 1              e - 1
    >  ( --------- )^x ~= N * -----
   /         N                  e
  '---
  x = 1

Note: I'm not sure how to demonstrate this but at least this is easily
verified with a simple program, the sum equals N * 0.632120 for any N
moderately large (tens to hundreds).

Inserting a constant sample value V here simply results in :

   sum = V * N * (e - 1) / e

But we don't want to integrate over a small period, but infinitely. Let's
cut the infinity in P periods of N values. Each period M is exactly the same
as period M-1 with a factor of ((N-1)/N)^N applied. A test shows that given a
large N :

     N - 1           1
  ( ------- )^N ~=  ---
       N             e

Our sum is now a sum of each factor times  :

   N*P                                     P
  ,---                                   ,---
   \         N - 1               e - 1    \     1
    >  v ( --------- )^x ~= VN * -----  *  >   ---
   /           N                   e      /    e^x
  '---                                   '---
  x = 1                                  x = 0

For P "large enough", in tests we get this :

    P
  ,---
   \     1        e
    >   --- ~=  -----
   /    e^x     e - 1
  '---
  x = 0

This simplifies the sum above :

   N*P
  ,---
   \         N - 1
    >  v ( --------- )^x = VN
   /           N
  '---
  x = 1

So basically by summing values and applying the last result an (N-1)/N factor
we just get N times the values over the long term, so we can recover the
constant value V by dividing by N.

A value added at the entry of the sliding window of N values will thus be
reduced to 1/e or 36.7% after N terms have been added. After a second batch,
it will only be 1/e^2, or 13.5%, and so on. So practically speaking, each
old period of N values represents only a quickly fading ratio of the global
sum :

  period    ratio
    1       36.7%
    2       13.5%
    3       4.98%
    4       1.83%
    5       0.67%
    6       0.25%
    7       0.09%
    8       0.033%
    9       0.012%
   10       0.0045%

So after 10N samples, the initial value has already faded out by a factor of
22026, which is quite fast. If the sliding window is 1024 samples wide, it
means that a sample will only count for 1/22k of its initial value after 10k
samples went after it, which results in half of the value it would represent
using an arithmetic mean. The benefit of this method is that it's very cheap
in terms of computations when N is a power of two. This is very well suited
to record response times as large values will fade out faster than with an
arithmetic mean and will depend on sample count and not time.

Demonstrating all the above assumptions with maths instead of a program is
left as an exercise for the reader.

diff --git a/include/proto/freq_ctr.h b/include/proto/freq_ctr.h
index d85e7210ff2b910f46195f8fc6ea4c1b3debeb37..65388b1f0d7f1201460d3b1192f147b43ac1e297 100644 (file)
--- a/include/proto/freq_ctr.h
+++ b/include/proto/freq_ctr.h
@@ -1,23 +1,23 @@
 /*
-  include/proto/freq_ctr.h
-  This file contains macros and inline functions for frequency counters.
-
-  Copyright (C) 2000-2009 Willy Tarreau - w@1wt.eu
-  
-  This library is free software; you can redistribute it and/or
-  modify it under the terms of the GNU Lesser General Public
-  License as published by the Free Software Foundation, version 2.1
-  exclusively.
-
-  This library 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
-  Lesser General Public License for more details.
-
-  You should have received a copy of the GNU Lesser General Public
-  License along with this library; if not, write to the Free Software
-  Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
-*/
+ * include/proto/freq_ctr.h
+ * This file contains macros and inline functions for frequency counters.
+ *
+ * Copyright (C) 2000-2014 Willy Tarreau - w@1wt.eu
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation, version 2.1
+ * exclusively.
+ *
+ * This library 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
+ */
 
 #ifndef _PROTO_FREQ_CTR_H
 #define _PROTO_FREQ_CTR_H
@@ -115,6 +115,123 @@ unsigned int read_freq_ctr_period(struct freq_ctr_period *ctr, unsigned int peri
 unsigned int freq_ctr_remain_period(struct freq_ctr_period *ctr, unsigned int period,
                                    unsigned int freq, unsigned int pend);
 
+/* While the functions above report average event counts per period, we are
+ * also interested in average values per event. For this we use a different
+ * method. The principle is to rely on a long tail which sums the new value
+ * with a fraction of the previous value, resulting in a sliding window of
+ * infinite length depending on the precision we're interested in.
+ *
+ * The idea is that we always keep (N-1)/N of the sum and add the new sampled
+ * value. The sum over N values can be computed with a simple program for a
+ * constant value 1 at each iteration :
+ *
+ *     N
+ *   ,---
+ *    \       N - 1              e - 1
+ *     >  ( --------- )^x ~= N * -----
+ *    /         N                  e
+ *   '---
+ *   x = 1
+ *
+ * Note: I'm not sure how to demonstrate this but at least this is easily
+ * verified with a simple program, the sum equals N * 0.632120 for any N
+ * moderately large (tens to hundreds).
+ *
+ * Inserting a constant sample value V here simply results in :
+ *
+ *    sum = V * N * (e - 1) / e
+ *
+ * But we don't want to integrate over a small period, but infinitely. Let's
+ * cut the infinity in P periods of N values. Each period M is exactly the same
+ * as period M-1 with a factor of ((N-1)/N)^N applied. A test shows that given a
+ * large N :
+ *
+ *      N - 1           1
+ *   ( ------- )^N ~=  ---
+ *        N             e
+ *
+ * Our sum is now a sum of each factor times  :
+ *
+ *    N*P                                     P
+ *   ,---                                   ,---
+ *    \         N - 1               e - 1    \     1
+ *     >  v ( --------- )^x ~= VN * -----  *  >   ---
+ *    /           N                   e      /    e^x
+ *   '---                                   '---
+ *   x = 1                                  x = 0
+ *
+ * For P "large enough", in tests we get this :
+ *
+ *    P
+ *  ,---
+ *   \     1        e
+ *    >   --- ~=  -----
+ *   /    e^x     e - 1
+ *  '---
+ *  x = 0
+ *
+ * This simplifies the sum above :
+ *
+ *    N*P
+ *   ,---
+ *    \         N - 1
+ *     >  v ( --------- )^x = VN
+ *    /           N
+ *   '---
+ *   x = 1
+ *
+ * So basically by summing values and applying the last result an (N-1)/N factor
+ * we just get N times the values over the long term, so we can recover the
+ * constant value V by dividing by N.
+ *
+ * A value added at the entry of the sliding window of N values will thus be
+ * reduced to 1/e or 36.7% after N terms have been added. After a second batch,
+ * it will only be 1/e^2, or 13.5%, and so on. So practically speaking, each
+ * old period of N values represents only a quickly fading ratio of the global
+ * sum :
+ *
+ *   period    ratio
+ *     1       36.7%
+ *     2       13.5%
+ *     3       4.98%
+ *     4       1.83%
+ *     5       0.67%
+ *     6       0.25%
+ *     7       0.09%
+ *     8       0.033%
+ *     9       0.012%
+ *    10       0.0045%
+ *
+ * So after 10N samples, the initial value has already faded out by a factor of
+ * 22026, which is quite fast. If the sliding window is 1024 samples wide, it
+ * means that a sample will only count for 1/22k of its initial value after 10k
+ * samples went after it, which results in half of the value it would represent
+ * using an arithmetic mean. The benefit of this method is that it's very cheap
+ * in terms of computations when N is a power of two. This is very well suited
+ * to record response times as large values will fade out faster than with an
+ * arithmetic mean and will depend on sample count and not time.
+ *
+ * Demonstrating all the above assumptions with maths instead of a program is
+ * left as an exercise for the reader.
+ */
+
+/* Adds sample value <v> to sliding window sum <sum> configured for <n> samples.
+ * The sample is returned. Better if <n> is a power of two.
+ */
+static inline unsigned int swrate_add(unsigned int *sum, unsigned int n, unsigned int v)
+{
+       return *sum = *sum * (n - 1) / n + v;
+}
+
+/* Returns the average sample value for the sum <sum> over a sliding window of
+ * <n> samples. Better if <n> is a power of two. It must be the same <n> as the
+ * one used above in all additions.
+ */
+static inline unsigned int swrate_avg(unsigned int sum, unsigned int n)
+{
+       return (sum + n - 1) / n;
+}
+
 #endif /* _PROTO_FREQ_CTR_H */
 
 /*