skia2/bench/gen_bench_expectations.py

224 lines
8.3 KiB
Python
Raw Normal View History

#!/usr/bin/env python
# Copyright (c) 2014 The Chromium Authors. All rights reserved.
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.
""" Generate bench_expectations file from a given set of bench data files. """
import argparse
import bench_util
import json
import os
import re
import sys
import urllib2
# Parameters for calculating bench ranges.
RANGE_RATIO_UPPER = 1.5 # Ratio of range for upper bounds.
RANGE_RATIO_LOWER = 2.0 # Ratio of range for lower bounds.
ERR_RATIO = 0.08 # Further widens the range by the ratio of average value.
ERR_UB = 1.0 # Adds an absolute upper error to cope with small benches.
ERR_LB = 1.5
# List of bench configs to monitor. Ignore all other configs.
CONFIGS_TO_INCLUDE = ['simple_viewport_1000x1000',
'simple_viewport_1000x1000_angle',
'simple_viewport_1000x1000_gpu',
'simple_viewport_1000x1000_scalar_1.100000',
'simple_viewport_1000x1000_scalar_1.100000_gpu',
]
# List of flaky entries that should be excluded. Each entry is defined by a list
# of 3 strings, corresponding to the substrings of [bench, config, builder] to
# search for. A bench expectations line is excluded when each of the 3 strings
# in the list is a substring of the corresponding element of the given line. For
# instance, ['desk_yahooanswers', 'gpu', 'Ubuntu'] will skip expectation entries
# of SKP benchs whose name contains 'desk_yahooanswers' on all gpu-related
# configs of all Ubuntu builders.
ENTRIES_TO_EXCLUDE = [
]
_GS_CLOUD_FORMAT = 'http://storage.googleapis.com/chromium-skia-gm/perfdata/%s/%s'
def compute_ranges(benches, more_benches=None):
"""Given a list of bench numbers, calculate the alert range.
Args:
benches: a list of float bench values.
more_benches: a tuple of lists of additional bench values.
The first value of each tuple is the number of commits before the current
one that set of values is at, and the second value is a list of
bench results.
Returns:
a list of float [lower_bound, upper_bound].
"""
avg = sum(benches)/len(benches)
minimum = min(benches)
maximum = max(benches)
diff = maximum - minimum
return [minimum - diff*RANGE_RATIO_LOWER - avg*ERR_RATIO - ERR_LB,
maximum + diff*RANGE_RATIO_UPPER + avg*ERR_RATIO + ERR_UB]
def create_expectations_dict(revision_data_points, builder, extra_data=None):
"""Convert list of bench data points into a dictionary of expectations data.
Args:
revision_data_points: a list of BenchDataPoint objects.
builder: string of the corresponding buildbot builder name.
Returns:
a dictionary of this form:
keys = tuple of (config, bench) strings.
values = list of float [expected, lower_bound, upper_bound] for the key.
"""
bench_dict = {}
for point in revision_data_points:
if (point.time_type or # Not walltime which has time_type ''
not point.config in CONFIGS_TO_INCLUDE):
continue
to_skip = False
for bench_substr, config_substr, builder_substr in ENTRIES_TO_EXCLUDE:
if (bench_substr in point.bench and config_substr in point.config and
builder_substr in builder):
to_skip = True
break
if to_skip:
continue
key = (point.config, point.bench)
extras = []
for idx, dataset in extra_data:
for data in dataset:
if (data.bench == point.bench and data.config == point.config and
data.time_type == point.time_type and data.per_iter_time):
extras.append((idx, data.per_iter_time))
if key in bench_dict:
raise Exception('Duplicate bench entry: ' + str(key))
bench_dict[key] = [point.time] + compute_ranges(point.per_iter_time, extras)
return bench_dict
def get_parent_commits(start_hash, num_back):
"""Returns a list of commits that are the parent of the commit passed in."""
list_commits = urllib2.urlopen(
'https://skia.googlesource.com/skia/+log/%s?format=json&n=%d' %
(start_hash, num_back))
# NOTE: Very brittle. Removes the four extraneous characters
# so json can be read successfully
trunc_list = list_commits.read()[4:]
json_data = json.loads(trunc_list)
return [revision['commit'] for revision in json_data['log']]
def get_file_suffixes(commit_hash, directory):
"""Gets all the suffixes available in the directory"""
possible_files = os.listdir(directory)
prefix = 'bench_' + commit_hash + '_data_'
return [name[len(prefix):] for name in possible_files
if name.startswith(prefix)]
def download_bench_data(builder, commit_hash, suffixes, directory):
"""Downloads data, returns the number successfully downloaded"""
cur_files = os.listdir(directory)
count = 0
for suffix in suffixes:
file_name = 'bench_'+commit_hash+'_data_'+suffix
if file_name in cur_files:
continue
try:
src = urllib2.urlopen(_GS_CLOUD_FORMAT % (builder, file_name))
with open(os.path.join(directory, file_name), 'w') as dest:
dest.writelines(src)
count += 1
except urllib2.HTTPError:
pass
return count
def main():
"""Reads bench data points, then calculate and export expectations.
"""
parser = argparse.ArgumentParser()
parser.add_argument(
'-a', '--representation_alg', default='25th',
help='bench representation algorithm to use, see bench_util.py.')
parser.add_argument(
'-b', '--builder', required=True,
help='name of the builder whose bench ranges we are computing.')
parser.add_argument(
'-d', '--input_dir', required=True,
help='a directory containing bench data files.')
parser.add_argument(
'-o', '--output_file', required=True,
help='file path and name for storing the output bench expectations.')
parser.add_argument(
'-r', '--git_revision', required=True,
help='the git hash to indicate the revision of input data to use.')
parser.add_argument(
'-t', '--back_track', required=False, default=10,
help='the number of commit hashes backwards to look to include' +
'in the calculations.')
parser.add_argument(
'-m', '--max_commits', required=False, default=1,
help='the number of commit hashes to include in the calculations.')
args = parser.parse_args()
builder = args.builder
data_points = bench_util.parse_skp_bench_data(
args.input_dir, args.git_revision, args.representation_alg)
parent_commits = get_parent_commits(args.git_revision, args.back_track)
print "Using commits: {}".format(parent_commits)
suffixes = get_file_suffixes(args.git_revision, args.input_dir)
print "Using suffixes: {}".format(suffixes)
# TODO(kelvinly): Find a better approach to than directly copying from
# the GS server?
downloaded_commits = []
for idx, commit in enumerate(parent_commits):
num_downloaded = download_bench_data(
builder, commit, suffixes, args.input_dir)
if num_downloaded > 0:
downloaded_commits.append((num_downloaded, idx, commit))
if len(downloaded_commits) < args.max_commits:
print ('Less than desired number of commits found. Please increase'
'--back_track in later runs')
trunc_commits = sorted(downloaded_commits, reverse=True)[:args.max_commits]
extra_data = []
for _, idx, commit in trunc_commits:
extra_data.append((idx, bench_util.parse_skp_bench_data(
args.input_dir, commit, args.representation_alg)))
expectations_dict = create_expectations_dict(data_points, builder,
extra_data)
out_lines = []
keys = expectations_dict.keys()
keys.sort()
for (config, bench) in keys:
(expected, lower_bound, upper_bound) = expectations_dict[(config, bench)]
out_lines.append('%(bench)s_%(config)s_,%(builder)s-%(representation)s,'
'%(expected)s,%(lower_bound)s,%(upper_bound)s' % {
'bench': bench,
'config': config,
'builder': builder,
'representation': args.representation_alg,
'expected': expected,
'lower_bound': lower_bound,
'upper_bound': upper_bound})
with open(args.output_file, 'w') as file_handle:
file_handle.write('\n'.join(out_lines))
if __name__ == "__main__":
main()