benchtests/scripts/compare_bench.py

   1 #!/usr/bin/python
   2 # Copyright (C) 2015-2018 Free Software Foundation, Inc.
   3 # This file is part of the GNU C Library.
   4 #
   5 # The GNU C Library is free software; you can redistribute it and/or
   6 # modify it under the terms of the GNU Lesser General Public
   7 # License as published by the Free Software Foundation; either
   8 # version 2.1 of the License, or (at your option) any later version.
   9 #
  10 # The GNU C Library is distributed in the hope that it will be useful,
  11 # but WITHOUT ANY WARRANTY; without even the implied warranty of
  12 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13 # Lesser General Public License for more details.
  14 #
  15 # You should have received a copy of the GNU Lesser General Public
  16 # License along with the GNU C Library; if not, see
  17 # <http://www.gnu.org/licenses/>.
  18 """Compare two benchmark results
  19
  20 Given two benchmark result files and a threshold, this script compares the
  21 benchmark results and flags differences in performance beyond a given
  22 threshold.
  23 """
  24 import sys
  25 import os
  26 import pylab
  27 import import_bench as bench
  28 import argparse
  29
  30 def do_compare(func, var, tl1, tl2, par, threshold):
  31     """Compare one of the aggregate measurements
  32
  33     Helper function to compare one of the aggregate measurements of a function
  34     variant.
  35
  36     Args:
  37         func: Function name
  38         var: Function variant name
  39         tl1: The first timings list
  40         tl2: The second timings list
  41         par: The aggregate to measure
  42         threshold: The threshold for differences, beyond which the script should
  43         print a warning.
  44     """
  45     d = abs(tl2[par] - tl1[par]) * 100 / tl1[str(par)]
  46     if d > threshold:
  47         if tl1[par] > tl2[par]:
  48             ind = '+++'
  49         else:
  50             ind = '---'
  51         print('%s %s(%s)[%s]: (%.2lf%%) from %g to %g' %
  52                 (ind, func, var, par, d, tl1[par], tl2[par]))
  53
  54
  55 def compare_runs(pts1, pts2, threshold):
  56     """Compare two benchmark runs
  57
  58     Args:
  59         pts1: Timing data from first machine
  60         pts2: Timing data from second machine
  61     """
  62
  63     # XXX We assume that the two benchmarks have identical functions and
  64     # variants.  We cannot compare two benchmarks that may have different
  65     # functions or variants.  Maybe that is something for the future.
  66     for func in pts1['functions'].keys():
  67         for var in pts1['functions'][func].keys():
  68             tl1 = pts1['functions'][func][var]
  69             tl2 = pts2['functions'][func][var]
  70
  71             # Compare the consolidated numbers
  72             # do_compare(func, var, tl1, tl2, 'max', threshold)
  73             do_compare(func, var, tl1, tl2, 'min', threshold)
  74             do_compare(func, var, tl1, tl2, 'mean', threshold)
  75
  76             # Skip over to the next variant or function if there is no detailed
  77             # timing info for the function variant.
  78             if 'timings' not in pts1['functions'][func][var].keys() or \
  79                 'timings' not in pts2['functions'][func][var].keys():
  80                     return
  81
  82             # If two lists do not have the same length then it is likely that
  83             # the performance characteristics of the function have changed.
  84             # XXX: It is also likely that there was some measurement that
  85             # strayed outside the usual range.  Such ouiers should not
  86             # happen on an idle machine with identical hardware and
  87             # configuration, but ideal environments are hard to come by.
  88             if len(tl1['timings']) != len(tl2['timings']):
  89                 print('* %s(%s): Timing characteristics changed' %
  90                         (func, var))
  91                 print('\tBefore: [%s]' %
  92                         ', '.join([str(x) for x in tl1['timings']]))
  93                 print('\tAfter: [%s]' %
  94                         ', '.join([str(x) for x in tl2['timings']]))
  95                 continue
  96
  97             # Collect numbers whose differences cross the threshold we have
  98             # set.
  99             issues = [(x, y) for x, y in zip(tl1['timings'], tl2['timings']) \
 100                         if abs(y - x) * 100 / x > threshold]
 101
 102             # Now print them.
 103             for t1, t2 in issues:
 104                 d = abs(t2 - t1) * 100 / t1
 105                 if t2 > t1:
 106                     ind = '-'
 107                 else:
 108                     ind = '+'
 109
 110                 print("%s %s(%s): (%.2lf%%) from %g to %g" %
 111                         (ind, func, var, d, t1, t2))
 112
 113
 114 def plot_graphs(bench1, bench2):
 115     """Plot graphs for functions
 116
 117     Make scatter plots for the functions and their variants.
 118
 119     Args:
 120         bench1: Set of points from the first machine
 121         bench2: Set of points from the second machine.
 122     """
 123     for func in bench1['functions'].keys():
 124         for var in bench1['functions'][func].keys():
 125             # No point trying to print a graph if there are no detailed
 126             # timings.
 127             if u'timings' not in bench1['functions'][func][var].keys():
 128                 print('Skipping graph for %s(%s)' % (func, var))
 129                 continue
 130
 131             pylab.clf()
 132             pylab.ylabel('Time (cycles)')
 133
 134             # First set of points
 135             length = len(bench1['functions'][func][var]['timings'])
 136             X = [float(x) for x in range(length)]
 137             lines = pylab.scatter(X, bench1['functions'][func][var]['timings'],
 138                     1.5 + 100 / length)
 139             pylab.setp(lines, 'color', 'r')
 140
 141             # Second set of points
 142             length = len(bench2['functions'][func][var]['timings'])
 143             X = [float(x) for x in range(length)]
 144             lines = pylab.scatter(X, bench2['functions'][func][var]['timings'],
 145                     1.5 + 100 / length)
 146             pylab.setp(lines, 'color', 'g')
 147
 148             if var:
 149                 filename = "%s-%s.png" % (func, var)
 150             else:
 151                 filename = "%s.png" % func
 152             print('Writing out %s' % filename)
 153             pylab.savefig(filename)
 154
 155 def main(bench1, bench2, schema, threshold):
 156     bench1 = bench.parse_bench(bench1, schema)
 157     bench2 = bench.parse_bench(bench2, schema)
 158
 159     plot_graphs(bench1, bench2)
 160
 161     bench.compress_timings(bench1)
 162     bench.compress_timings(bench2)
 163
 164     compare_runs(bench1, bench2, threshold)
 165
 166
 167 if __name__ == '__main__':
 168     parser = argparse.ArgumentParser(description='Take two benchmark and compare their timings.')
 169
 170     # Required parameters
 171     parser.add_argument('bench1', help='First bench to compare')
 172     parser.add_argument('bench2', help='Second bench to compare')
 173
 174     # Optional parameters
 175     parser.add_argument('--schema',
 176                         default=os.path.join(os.path.dirname(os.path.realpath(__file__)),'benchout.schema.json'),
 177                         help='JSON file to validate source/dest files (default: %(default)s)')
 178     parser.add_argument('--threshold', default=10.0, help='Only print those with equal or higher threshold (default: %(default)s)')
 179
 180     args = parser.parse_args()
 181
 182     main(args.bench1, args.bench2, args.schema, args.threshold)