benchtest: script to compare two benchmarks

This script is a sample implementation that uses import_bench to
construct two benchmark objects and compare them.  If detailed timing
information is available (when one does `make DETAILED=1 bench`), it
writes out graphs for all functions it benchmarks and prints
significant differences in timings of the two benchmark runs.  If
detailed timing information is not available, it points out
significant differences in aggregate times.

Call this script as follows:

  compare_bench.py schema_file.json bench1.out bench2.out

Alternatively, if one wants to set a different threshold for warnings
(default is a 10% difference):

  compare_bench.py schema_file.json bench1.out bench2.out 25

The threshold in the example above is 25%.  schema_file.json is the
JSON schema (which is $srcdir/benchtests/scripts/benchout.schema.json
for the benchmark output file) and bench1.out and bench2.out are the
two benchmark output files to compare.

The key functionality here is the compress_timings function which
groups together points that are close together into a single point
that is the mean of all its representative points.  Any point in such
a group is at most 1.5x the smallest point in that group.  The
detailed derivation is a comment in the function.

	* benchtests/scripts/compare_bench.py: New file.
	* benchtests/scripts/import_bench.py (mean): New function.
	(split_list): Likewise.
	(do_for_all_timings): Likewise.
	(compress_timings): Likewise.

2 files changed, 280 insertions, 0 deletions

diff --git a/benchtests/scripts/compare_bench.py b/benchtests/scripts/compare_bench.py
new file mode 100755
index 0000000000..be5b5ca9cd
--- /dev/null
+++ b/benchtests/scripts/compare_bench.py
@@ -0,0 +1,184 @@
+#!/usr/bin/python
+# Copyright (C) 2015 Free Software Foundation, Inc.
+# This file is part of the GNU C Library.
+#
+# The GNU C 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; either
+# version 2.1 of the License, or (at your option) any later version.
+#
+# The GNU C 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 the GNU C Library; if not, see
+# <http://www.gnu.org/licenses/>.
+"""Compare two benchmark results
+
+Given two benchmark result files and a threshold, this script compares the
+benchmark results and flags differences in performance beyond a given
+threshold.
+"""
+import sys
+import os
+import pylab
+import import_bench as bench
+
+def do_compare(func, var, tl1, tl2, par, threshold):
+    """Compare one of the aggregate measurements
+
+    Helper function to compare one of the aggregate measurements of a function
+    variant.
+
+    Args:
+        func: Function name
+        var: Function variant name
+        tl1: The first timings list
+        tl2: The second timings list
+        par: The aggregate to measure
+        threshold: The threshold for differences, beyond which the script should
+        print a warning.
+    """
+    d = abs(tl2[par] - tl1[par]) * 100 / tl1[str(par)]
+    if d > threshold:
+        if tl1[par] > tl2[par]:
+            ind = '+++'
+        else:
+            ind = '---'
+        print('%s %s(%s)[%s]: (%.2lf%%) from %g to %g' %
+                (ind, func, var, par, d, tl1[par], tl2[par]))
+
+
+def compare_runs(pts1, pts2, threshold):
+    """Compare two benchmark runs
+
+    Args:
+        pts1: Timing data from first machine
+        pts2: Timing data from second machine
+    """
+
+    # XXX We assume that the two benchmarks have identical functions and
+    # variants.  We cannot compare two benchmarks that may have different
+    # functions or variants.  Maybe that is something for the future.
+    for func in pts1['functions'].keys():
+        for var in pts1['functions'][func].keys():
+            tl1 = pts1['functions'][func][var]
+            tl2 = pts2['functions'][func][var]
+
+            # Compare the consolidated numbers
+            # do_compare(func, var, tl1, tl2, 'max', threshold)
+            do_compare(func, var, tl1, tl2, 'min', threshold)
+            do_compare(func, var, tl1, tl2, 'mean', threshold)
+
+            # Skip over to the next variant or function if there is no detailed
+            # timing info for the function variant.
+            if 'timings' not in pts1['functions'][func][var].keys() or \
+                'timings' not in pts2['functions'][func][var].keys():
+                    return
+
+            # If two lists do not have the same length then it is likely that
+            # the performance characteristics of the function have changed.
+            # XXX: It is also likely that there was some measurement that
+            # strayed outside the usual range.  Such ouiers should not
+            # happen on an idle machine with identical hardware and
+            # configuration, but ideal environments are hard to come by.
+            if len(tl1['timings']) != len(tl2['timings']):
+                print('* %s(%s): Timing characteristics changed' %
+                        (func, var))
+                print('\tBefore: [%s]' %
+                        ', '.join([str(x) for x in tl1['timings']]))
+                print('\tAfter: [%s]' %
+                        ', '.join([str(x) for x in tl2['timings']]))
+                continue
+
+            # Collect numbers whose differences cross the threshold we have
+            # set.
+            issues = [(x, y) for x, y in zip(tl1['timings'], tl2['timings']) \
+                        if abs(y - x) * 100 / x > threshold]
+
+            # Now print them.
+            for t1, t2 in issues:
+                d = abs(t2 - t1) * 100 / t1
+                if t2 > t1:
+                    ind = '-'
+                else:
+                    ind = '+'
+
+                print("%s %s(%s): (%.2lf%%) from %g to %g" %
+                        (ind, func, var, d, t1, t2))
+
+
+def plot_graphs(bench1, bench2):
+    """Plot graphs for functions
+
+    Make scatter plots for the functions and their variants.
+
+    Args:
+        bench1: Set of points from the first machine
+        bench2: Set of points from the second machine.
+    """
+    for func in bench1['functions'].keys():
+        for var in bench1['functions'][func].keys():
+            # No point trying to print a graph if there are no detailed
+            # timings.
+            if u'timings' not in bench1['functions'][func][var].keys():
+                print('Skipping graph for %s(%s)' % (func, var))
+                continue
+
+            pylab.clf()
+            pylab.ylabel('Time (cycles)')
+
+            # First set of points
+            length = len(bench1['functions'][func][var]['timings'])
+            X = [float(x) for x in range(length)]
+            lines = pylab.scatter(X, bench1['functions'][func][var]['timings'],
+                    1.5 + 100 / length)
+            pylab.setp(lines, 'color', 'r')
+
+            # Second set of points
+            length = len(bench2['functions'][func][var]['timings'])
+            X = [float(x) for x in range(length)]
+            lines = pylab.scatter(X, bench2['functions'][func][var]['timings'],
+                    1.5 + 100 / length)
+            pylab.setp(lines, 'color', 'g')
+
+            if var:
+                filename = "%s-%s.png" % (func, var)
+            else:
+                filename = "%s.png" % func
+            print('Writing out %s' % filename)
+            pylab.savefig(filename)
+
+
+def main(args):
+    """Program Entry Point
+
+    Take two benchmark output files and compare their timings.
+    """
+    if len(args) > 4 or len(args) < 3:
+        print('Usage: %s <schema> <file1> <file2> [threshold in %%]' % sys.argv[0])
+        sys.exit(os.EX_USAGE)
+
+    bench1 = bench.parse_bench(args[1], args[0])
+    bench2 = bench.parse_bench(args[2], args[0])
+    if len(args) == 4:
+        threshold = float(args[3])
+    else:
+        threshold = 10.0
+
+    if (bench1['timing_type'] != bench2['timing_type']):
+        print('Cannot compare benchmark outputs: timing types are different')
+        return
+
+    plot_graphs(bench1, bench2)
+
+    bench.compress_timings(bench1)
+    bench.compress_timings(bench2)
+
+    compare_runs(bench1, bench2, threshold)
+
+
+if __name__ == '__main__':
+    main(sys.argv[1:])
diff --git a/benchtests/scripts/import_bench.py b/benchtests/scripts/import_bench.py
index 81248c2adf..d37ff62383 100644
--- a/benchtests/scripts/import_bench.py
+++ b/benchtests/scripts/import_bench.py
@@ -25,6 +25,102 @@ except ImportError:
     raise
 
 
+def mean(lst):
+    """Compute and return mean of numbers in a list
+
+    The numpy average function has horrible performance, so implement our
+    own mean function.
+
+    Args:
+        lst: The list of numbers to average.
+    Return:
+        The mean of members in the list.
+    """
+    return sum(lst) / len(lst)
+
+
+def split_list(bench, func, var):
+    """ Split the list into a smaller set of more distinct points
+
+    Group together points such that the difference between the smallest
+    point and the mean is less than 1/3rd of the mean.  This means that
+    the mean is at most 1.5x the smallest member of that group.
+
+    mean - xmin < mean / 3
+    i.e. 2 * mean / 3 < xmin
+    i.e. mean < 3 * xmin / 2
+
+    For an evenly distributed group, the largest member will be less than
+    twice the smallest member of the group.
+    Derivation:
+
+    An evenly distributed series would be xmin, xmin + d, xmin + 2d...
+
+    mean = (2 * n * xmin + n * (n - 1) * d) / 2 * n
+    and max element is xmin + (n - 1) * d
+
+    Now, mean < 3 * xmin / 2
+
+    3 * xmin > 2 * mean
+    3 * xmin > (2 * n * xmin + n * (n - 1) * d) / n
+    3 * n * xmin > 2 * n * xmin + n * (n - 1) * d
+    n * xmin > n * (n - 1) * d
+    xmin > (n - 1) * d
+    2 * xmin > xmin + (n-1) * d
+    2 * xmin > xmax
+
+    Hence, proved.
+
+    Similarly, it is trivial to prove that for a similar aggregation by using
+    the maximum element, the maximum element in the group must be at most 4/3
+    times the mean.
+
+    Args:
+        bench: The benchmark object
+        func: The function name
+        var: The function variant name
+    """
+    means = []
+    lst = bench['functions'][func][var]['timings']
+    last = len(lst) - 1
+    while lst:
+        for i in range(last + 1):
+            avg = mean(lst[i:])
+            if avg > 0.75 * lst[last]:
+                means.insert(0, avg)
+                lst = lst[:i]
+                last = i - 1
+                break
+    bench['functions'][func][var]['timings'] = means
+
+
+def do_for_all_timings(bench, callback):
+    """Call a function for all timing objects for each function and its
+    variants.
+
+    Args:
+        bench: The benchmark object
+        callback: The callback function
+    """
+    for func in bench['functions'].keys():
+        for k in bench['functions'][func].keys():
+            if 'timings' not in bench['functions'][func][k].keys():
+                continue
+
+            callback(bench, func, k)
+
+
+def compress_timings(points):
+    """Club points with close enough values into a single mean value
+
+    See split_list for details on how the clubbing is done.
+
+    Args:
+        points: The set of points.
+    """
+    do_for_all_timings(points, split_list)
+
+
 def parse_bench(filename, schema_filename):
     """Parse the input file