OpenE2K
/
glibc
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

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.
This commit is contained in:
Siddhesh Poyarekar 2015-06-01 23:14:11 +05:30
parent 0994b9b6f6
commit 0cd2828695
3 changed files with 286 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ChangeLog
View File

@ -1,5 +1,11 @@
2015-06-01 Siddhesh Poyarekar <siddhesh@redhat.com>
* 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.
* benchtests/scripts/import_bench.py: New file.
* benchtests/scripts/validate_benchout.py: Import import_bench
instead of jsonschema.

184
benchtests/scripts/compare_bench.py Executable file
View File

@ -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:])

96
benchtests/scripts/import_bench.py
View File

@ -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