path: root/chromium/testing/scripts/run_performance_tests.py

#!/usr/bin/env vpython3
# Copyright 2017 The Chromium Authors
# Use of this source code is governed by a BSD-style license that can be
# found in the LICENSE file.

"""Runs telemetry benchmarks and gtest perf tests.

If optional argument --isolated-script-test-output=[FILENAME] is passed
to the script, json is written to that file in the format detailed in
//docs/testing/json-test-results-format.md.

If optional argument --isolated-script-test-filter=[TEST_NAMES] is passed to
the script, it should be a  double-colon-separated ("::") list of test names,
to run just that subset of tests.

This script is intended to be the base command invoked by the isolate,
followed by a subsequent Python script. It could be generalized to
invoke an arbitrary executable.
It currently runs several benchmarks. The benchmarks it will execute are
based on the shard it is running on and the sharding_map_path.

If this is executed with a gtest perf test, the flag --non-telemetry
has to be passed in to the script so the script knows it is running
an executable and not the run_benchmark command.

This script merges test results from all the benchmarks into the one
output.json file. The test results and perf results are also put in separate
directories per benchmark. Two files will be present in each directory;
perf_results.json, which is the perf specific results (with unenforced format,
could be histogram or graph json), and test_results.json.

TESTING:
To test changes to this script, please run
cd tools/perf
./run_tests ScriptsSmokeTest.testRunPerformanceTests
"""

from __future__ import print_function

import argparse
import json
import os
import requests
import shutil
import sys
import time
import tempfile
import traceback
import six

from collections import OrderedDict

CHROMIUM_SRC_DIR = os.path.abspath(
    os.path.join(os.path.dirname(__file__),
                 os.path.pardir, os.path.pardir))

PERF_DIR = os.path.join(CHROMIUM_SRC_DIR, 'tools', 'perf')
sys.path.append(PERF_DIR)
import generate_legacy_perf_dashboard_json
from core import path_util

PERF_CORE_DIR = os.path.join(PERF_DIR, 'core')
sys.path.append(PERF_CORE_DIR)
import results_merger

# Add src/testing/ into sys.path for importing xvfb, test_env, and common.
sys.path.append(
    os.path.abspath(os.path.join(os.path.dirname(__file__), os.path.pardir)))
import xvfb
import test_env
from scripts import common

SHARD_MAPS_DIRECTORY = os.path.abspath(
    os.path.join(os.path.dirname(__file__), os.path.pardir, os.path.pardir,
                 'tools', 'perf', 'core', 'shard_maps'))

# See https://crbug.com/923564.
# We want to switch over to using histograms for everything, but converting from
# the format output by gtest perf tests to histograms has introduced several
# problems. So, only perform the conversion on tests that are whitelisted and
# are okay with potentially encountering issues.
GTEST_CONVERSION_WHITELIST = [
  'angle_perftests',
  'base_perftests',
  'blink_heap_perftests',
  'blink_platform_perftests',
  'cc_perftests',
  'components_perftests',
  'command_buffer_perftests',
  'dawn_perf_tests',
  'gpu_perftests',
  'load_library_perf_tests',
  'net_perftests',
  'browser_tests',
  'services_perftests',
  # TODO(jmadill): Remove once migrated. http://anglebug.com/5124
  'standalone_angle_perftests',
  'sync_performance_tests',
  'tracing_perftests',
  'views_perftests',
  'viz_perftests',
  'wayland_client_perftests',
  'xr.vr.common_perftests',
]

# pylint: disable=useless-object-inheritance


class OutputFilePaths(object):
  """Provide paths to where results outputs should be written.

  The process_perf_results.py merge script later will pull all of these
  together, so that's why they aren't in the standard locations. Also,
  note that because of the OBBS (One Build Bot Step), Telemetry
  has multiple tests running on a single shard, so we need to prefix
  these locations with a directory named by the benchmark name.
  """

  def __init__(self, isolated_out_dir, perf_test_name):
    self.name = perf_test_name
    self.benchmark_path = os.path.join(isolated_out_dir, perf_test_name)

  def SetUp(self):
    os.makedirs(self.benchmark_path)
    return self

  @property
  def perf_results(self):
    return os.path.join(self.benchmark_path, 'perf_results.json')

  @property
  def test_results(self):
    return os.path.join(self.benchmark_path, 'test_results.json')

  @property
  def logs(self):
    return os.path.join(self.benchmark_path, 'benchmark_log.txt')

  @property
  def csv_perf_results(self):
    """Path for csv perf results.

    Note that the chrome.perf waterfall uses the json histogram perf results
    exclusively. csv_perf_results are implemented here in case a user script
    passes --output-format=csv.
    """
    return os.path.join(self.benchmark_path, 'perf_results.csv')


def print_duration(step, start):
  print('Duration of %s: %d seconds' % (step, time.time() - start))


def IsWindows():
  return sys.platform == 'cygwin' or sys.platform.startswith('win')


class GtestCommandGenerator(object):
  def __init__(self, options, override_executable=None, additional_flags=None,
               ignore_shard_env_vars=False):
    self._options = options
    self._override_executable = override_executable
    self._additional_flags = additional_flags or []
    self._ignore_shard_env_vars = ignore_shard_env_vars

  def generate(self, output_dir):
    """Generate the command to run to start the gtest perf test.

    Returns:
      list of strings, the executable and its arguments.
    """
    return ([self._get_executable()] +
            self._generate_filter_args() +
            self._generate_repeat_args() +
            self._generate_also_run_disabled_tests_args() +
            self._generate_output_args(output_dir) +
            self._generate_shard_args() +
            self._get_additional_flags()
           )

  @property
  def executable_name(self):
    """Gets the platform-independent name of the executable."""
    return self._override_executable or self._options.executable

  def _get_executable(self):
    executable = str(self.executable_name)
    if IsWindows():
      return r'.\%s.exe' % executable
    return './%s' % executable

  def _get_additional_flags(self):
    return self._additional_flags

  def _generate_shard_args(self):
    """Teach the gtest to ignore the environment variables.

    GTEST_SHARD_INDEX and GTEST_TOTAL_SHARDS will confuse the gtest
    and convince it to only run some of its tests. Instead run all
    of them.
    """
    if self._ignore_shard_env_vars:
      return ['--test-launcher-total-shards=1', '--test-launcher-shard-index=0']
    return []

  def _generate_filter_args(self):
    if self._options.isolated_script_test_filter:
      filter_list = common.extract_filter_list(
        self._options.isolated_script_test_filter)
      return ['--gtest_filter=' + ':'.join(filter_list)]
    return []

  def _generate_repeat_args(self):
    # TODO(crbug.com/920002): Support --isolated-script-test-repeat.
    return []

  def _generate_also_run_disabled_tests_args(self):
    # TODO(crbug.com/920002): Support
    # --isolated-script-test-also-run-disabled-tests.
    return []

  def _generate_output_args(self, output_dir):
    output_args = []
    if self._options.use_gtest_benchmark_script:
      output_args.append('--output-dir=' + output_dir)
    # These flags are to make sure that test output perf metrics in the log.
    if not '--verbose' in self._get_additional_flags():
      output_args.append('--verbose')
    if (not '--test-launcher-print-test-stdio=always'
        in self._get_additional_flags()):
      output_args.append('--test-launcher-print-test-stdio=always')
    return output_args


def write_simple_test_results(return_code, output_filepath, benchmark_name):
  # TODO(crbug.com/1115658): Fix to output
  # https://chromium.googlesource.com/chromium/src/+/main/docs/testing/json_test_results_format.md
  # for each test rather than this summary.
  # Append the shard index to the end of the name so that the merge script
  # doesn't blow up trying to merge unmergeable results.
  benchmark_name += '_shard_%s' % os.environ.get('GTEST_SHARD_INDEX', '0')
  output_json = {
      'tests': {
          benchmark_name: {
              'expected': 'PASS',
              'actual': 'FAIL' if return_code else 'PASS',
              'is_unexpected': bool(return_code),
          },
      },
      'interrupted': False,
      'path_delimiter': '/',
      'version': 3,
      'seconds_since_epoch': time.time(),
      'num_failures_by_type': {
          'FAIL': 1 if return_code else 0,
          'PASS': 0 if return_code else 1,
      },
  }
  with open(output_filepath, 'w') as fh:
    json.dump(output_json, fh)


def upload_simple_test_results(return_code, benchmark_name):
  # TODO(crbug.com/1115658): Fix to upload results for each test rather than
  # this summary.
  try:
    with open(os.environ['LUCI_CONTEXT']) as f:
      sink = json.load(f)['result_sink']
  except KeyError:
    return

  if return_code:
    summary = '<p>Benchmark failed with status code %d</p>' % return_code
  else:
    summary = '<p>Benchmark passed</p>'

  result_json = {
      'testResults': [{
        'testId': benchmark_name,
        'expected': not return_code,
        'status': 'FAIL' if return_code else 'PASS',
        'summaryHtml': summary,
        'tags': [{'key': 'exit_code', 'value': str(return_code)}],
      }]
  }

  res = requests.post(
      url='http://%s/prpc/luci.resultsink.v1.Sink/ReportTestResults' %
      sink['address'],
      headers={
          'Content-Type': 'application/json',
          'Accept': 'application/json',
          'Authorization': 'ResultSink %s' % sink['auth_token'],
      },
      data=json.dumps(result_json))
  res.raise_for_status()


def execute_gtest_perf_test(command_generator, output_paths, use_xvfb=False,
                            is_unittest=False):
  start = time.time()

  env = os.environ.copy()
  env['CHROME_HEADLESS'] = '1'
  #TODO(crbug/1138988): Some gtests do not implements the unit_test_launcher.cc.
  # As a result, they will not respect the arguments added by
  # _generate_shard_args() and will still use the values of GTEST_SHARD_INDEX
  # and GTEST_TOTAL_SHARDS to run part of the tests.
  # Removing those environment variables as a workaround.
  if command_generator._ignore_shard_env_vars:
    if 'GTEST_TOTAL_SHARDS' in env:
      env.pop('GTEST_TOTAL_SHARDS')
    if 'GTEST_SHARD_INDEX' in env:
      env.pop('GTEST_SHARD_INDEX')

  return_code = 0
  try:
    command = command_generator.generate(output_paths.benchmark_path)
    if use_xvfb:
      # When running with xvfb, we currently output both to stdout and to the
      # file. It would be better to only output to the file to keep the logs
      # clean.
      return_code = xvfb.run_executable(
          command, env, stdoutfile=output_paths.logs)
    else:
      with open(output_paths.logs, 'w') as handle:
        try:
          return_code = test_env.run_command_output_to_handle(
              command, handle, env=env)
        except OSError as e:
          print('Command to run gtest perf test %s failed with an OSError: %s' %
                (output_paths.name, e))
          return_code = 1
    if (not os.path.exists(output_paths.perf_results) and
        os.path.exists(output_paths.logs)):
      # Get the correct json format from the stdout to write to the perf
      # results file if gtest does not generate one.
      results_processor = generate_legacy_perf_dashboard_json.\
          LegacyResultsProcessor()
      graph_json_string = results_processor.GenerateJsonResults(
          output_paths.logs)
      with open(output_paths.perf_results, 'w') as fh:
        fh.write(graph_json_string)
  except Exception:
    traceback.print_exc()
    return_code = 1
  if os.path.exists(output_paths.perf_results):
    if command_generator.executable_name in GTEST_CONVERSION_WHITELIST:
      with path_util.SysPath(path_util.GetTracingDir()):
        # pylint: disable=no-name-in-module,import-outside-toplevel
        from tracing.value import gtest_json_converter
        # pylint: enable=no-name-in-module,import-outside-toplevel
      gtest_json_converter.ConvertGtestJsonFile(output_paths.perf_results)
  else:
    print('ERROR: gtest perf test %s did not generate perf output' %
          output_paths.name)
    return_code = 1
  write_simple_test_results(return_code, output_paths.test_results,
                            output_paths.name)
  if not is_unittest:
    upload_simple_test_results(return_code, output_paths.name)

  print_duration(
      'executing gtest %s' % command_generator.executable_name, start)

  return return_code


class _TelemetryFilterArgument(object):
  def __init__(self, filter_string):
    self.benchmark, self.story = filter_string.split('/')


class TelemetryCommandGenerator(object):
  def __init__(self, benchmark, options,
               story_selection_config=None, is_reference=False):
    self.benchmark = benchmark
    self._options = options
    self._story_selection_config = story_selection_config
    self._is_reference = is_reference

  def generate(self, output_dir):
    """Generate the command to run to start the benchmark.

    Args:
      output_dir: The directory to configure the command to put output files
        into.

    Returns:
      list of strings, the executable and its arguments.
    """
    return ([sys.executable] + self._options.executable.split(' ') +
            [self.benchmark] +
            self._generate_filter_args() +
            self._generate_also_run_disabled_tests_args() +
            self._generate_output_args(output_dir) +
            self._generate_story_selection_args() +
            # passthrough args must be before reference args and repeat args:
            # crbug.com/928928, crbug.com/894254#c78
            self._get_passthrough_args() +
            self._generate_syslog_args() +
            self._generate_repeat_args() +
            self._generate_reference_build_args()
           )

  def _get_passthrough_args(self):
    return self._options.passthrough_args

  def _generate_filter_args(self):
    if self._options.isolated_script_test_filter:
      filter_list = common.extract_filter_list(
          self._options.isolated_script_test_filter)
      filter_arguments = [_TelemetryFilterArgument(f) for f in filter_list]
      applicable_stories = [
          f.story for f in filter_arguments if f.benchmark == self.benchmark]
      # Need to convert this to a valid regex.
      filter_regex = '(' + '|'.join(applicable_stories) + ')'
      return ['--story-filter=' + filter_regex]
    return []

  def _generate_repeat_args(self):
    if self._options.isolated_script_test_repeat:
      return ['--pageset-repeat=' + str(
          self._options.isolated_script_test_repeat)]
    return []

  def _generate_also_run_disabled_tests_args(self):
    if self._options.isolated_script_test_also_run_disabled_tests:
      return ['--also-run-disabled-tests']
    return []

  def _generate_output_args(self, output_dir):
    return ['--output-format=json-test-results',
            '--output-format=histograms',
            '--output-dir=' + output_dir]

  def _generate_story_selection_args(self):
    """Returns arguments that limit the stories to be run inside the benchmark.
    """
    selection_args = []
    if self._story_selection_config:
      if 'begin' in self._story_selection_config:
        selection_args.append('--story-shard-begin-index=%d' % (
            self._story_selection_config['begin']))
      if 'end' in self._story_selection_config:
        selection_args.append('--story-shard-end-index=%d' % (
            self._story_selection_config['end']))
      if 'sections' in self._story_selection_config:
        range_string = self._generate_story_index_ranges(
            self._story_selection_config['sections'])
        if range_string:
          selection_args.append('--story-shard-indexes=%s' % range_string)
      if self._story_selection_config.get('abridged', True):
        selection_args.append('--run-abridged-story-set')
    return selection_args

  def _generate_syslog_args(self):
    if self._options.per_test_logs_dir:
      isolated_out_dir = os.path.dirname(
          self._options.isolated_script_test_output)
      return ['--logs-dir', os.path.join(
          isolated_out_dir,
          self.benchmark)]
    return []


  def _generate_story_index_ranges(self, sections):
    range_string = ''
    for section in sections:
      begin = section.get('begin', '')
      end = section.get('end', '')
      # If there only one story in the range, we only keep its index.
      # In general, we expect either begin or end, or both.
      if begin != '' and end != '' and end - begin == 1:
        new_range = str(begin)
      elif begin != '' or end != '':
        new_range = '%s-%s' % (str(begin), str(end))
      else:
        raise ValueError('Index ranges in "sections" in shard map should have'
                         'at least one of "begin" and "end": %s' % str(section))
      if range_string:
        range_string += ',%s' % new_range
      else:
        range_string = new_range
    return range_string


  def _generate_reference_build_args(self):
    if self._is_reference:
      reference_browser_flag = '--browser=reference'
      # TODO(crbug.com/1038137): Make the logic generic once more reference
      # settings are added
      if '--browser=android-chrome-bundle' in self._get_passthrough_args():
        reference_browser_flag = '--browser=reference-android-chrome-bundle'
      return [reference_browser_flag,
              '--max-failures=5']
    return []


def execute_telemetry_benchmark(
    command_generator, output_paths, use_xvfb=False):
  start = time.time()

  env = os.environ.copy()
  env['CHROME_HEADLESS'] = '1'

  return_code = 1
  temp_dir = tempfile.mkdtemp('telemetry')
  try:
    command = command_generator.generate(temp_dir)
    if use_xvfb:
      # When running with xvfb, we currently output both to stdout and to the
      # file. It would be better to only output to the file to keep the logs
      # clean.
      return_code = xvfb.run_executable(
          command, env=env, stdoutfile=output_paths.logs)
    else:
      with open(output_paths.logs, 'w') as handle:
        return_code = test_env.run_command_output_to_handle(
            command, handle, env=env)
    expected_results_filename = os.path.join(temp_dir, 'test-results.json')
    if os.path.exists(expected_results_filename):
      shutil.move(expected_results_filename, output_paths.test_results)
    else:
      common.write_interrupted_test_results_to(output_paths.test_results, start)
    expected_perf_filename = os.path.join(temp_dir, 'histograms.json')
    shutil.move(expected_perf_filename, output_paths.perf_results)

    csv_file_path = os.path.join(temp_dir, 'results.csv')
    if os.path.isfile(csv_file_path):
      shutil.move(csv_file_path, output_paths.csv_perf_results)
  except Exception:
    print('The following exception may have prevented the code from '
          'outputing structured test results and perf results output:')
    print(traceback.format_exc())
  finally:
    # On swarming bots, don't remove output directory, since Result Sink might
    # still be uploading files to Result DB. Also, swarming bots automatically
    # clean up at the end of each task.
    if 'SWARMING_TASK_ID' not in os.environ:
      # Add ignore_errors=True because otherwise rmtree may fail due to leaky
      # processes of tests are still holding opened handles to files under
      # |tempfile_dir|. For example, see crbug.com/865896
      shutil.rmtree(temp_dir, ignore_errors=True)

  print_duration('executing benchmark %s' % command_generator.benchmark, start)

  # Telemetry sets exit code to -1 to indicate that no stories were run. This
  # becomes 255 on linux because linux doesn't support -1 so it does modulo:
  # -1 % 256 == 255.
  # TODO(crbug.com/1019139): Make 111 be the exit code that means
  # "no stories were run.".
  if return_code in (111, -1, 255):
    print('Exit code %s indicates that no stories were run, so we are marking '
          'this as a success.' % return_code)
    return 0
  if return_code:
    return return_code
  return 0

def parse_arguments(args):
  parser = argparse.ArgumentParser()
  parser.add_argument('executable', help='The name of the executable to run.')
  parser.add_argument(
      '--isolated-script-test-output', required=True)
  # The following two flags may be passed in sometimes by Pinpoint
  # or by the recipe, but they don't do anything. crbug.com/927482.
  parser.add_argument(
      '--isolated-script-test-chartjson-output', required=False)
  parser.add_argument(
      '--isolated-script-test-perf-output', required=False)

  parser.add_argument(
      '--isolated-script-test-filter', type=str, required=False)

  # Note that the following three arguments are only supported by Telemetry
  # tests right now. See crbug.com/920002.
  parser.add_argument(
      '--isolated-script-test-repeat', type=int, required=False)
  parser.add_argument(
      '--isolated-script-test-launcher-retry-limit', type=int, required=False,
      choices=[0])  # Telemetry does not support retries. crbug.com/894254#c21
  parser.add_argument(
      '--isolated-script-test-also-run-disabled-tests',
      default=False, action='store_true', required=False)
  parser.add_argument('--xvfb', help='Start xvfb.', action='store_true')
  parser.add_argument('--non-telemetry',
                      help='Type of perf test', type=bool, default=False)
  parser.add_argument('--gtest-benchmark-name',
                      help='Name of the gtest benchmark', type=str,
                      required=False)
  parser.add_argument('--use-gtest-benchmark-script',
                      help='Whether gtest is invoked via benchmark script.',
                      default=False, action='store_true')

  parser.add_argument('--benchmarks',
                      help='Comma separated list of benchmark names'
                      ' to run in lieu of indexing into our benchmark bot maps',
                      required=False)
  # crbug.com/1236245: This allows for per-benchmark device logs.
  parser.add_argument('--per-test-logs-dir',
                      help='Require --logs-dir args for test', required=False,
                      default=False, action='store_true')
  # Some executions may have a different sharding scheme and/or set of tests.
  # These files must live in src/tools/perf/core/shard_maps
  parser.add_argument('--test-shard-map-filename', type=str, required=False)
  parser.add_argument('--run-ref-build',
                      help='Run test on reference browser', action='store_true')
  parser.add_argument('--passthrough-arg',
                      help='Arguments to pass directly through to the test '
                      'executable.', action='append',
                      dest='passthrough_args',
                      default=[])
  parser.add_argument('--use-dynamic-shards',
                      help='If set, use dynamic shardmap instead of the file.',
                      action='store_true',
                      required=False
                      )
  parser.add_argument('--dynamic-shardmap',
                      help='The dynamically generated shardmap string used to '
                      'replace the static shardmap file.',
                      type=str,
                      required=False)
  options, leftover_args = parser.parse_known_args(args)
  options.passthrough_args.extend(leftover_args)
  return options


def main(sys_args):
  args = sys_args[1:]  # Skip program name.
  options = parse_arguments(args)
  isolated_out_dir = os.path.dirname(options.isolated_script_test_output)
  overall_return_code = 0
  # This is a list of test results files to be merged into a standard
  # output.json file for use by infrastructure including FindIt.
  # This list should not contain reference build runs
  # since we do not monitor those. Also, merging test reference build results
  # with standard build results may not work properly.
  test_results_files = []

  print('Running a series of performance test subprocesses. Logs, performance\n'
        'results, and test results JSON will be saved in a subfolder of the\n'
        'isolated output directory. Inside the hash marks in the following\n'
        'lines is the name of the subfolder to find results in.\n')

  if options.non_telemetry:
    benchmark_name = options.gtest_benchmark_name
    passthrough_args = options.passthrough_args
    # crbug/1146949#c15
    # In the case that pinpoint passes all arguments to swarming through http
    # request, the passthrough_args are converted into a comma-separated string.
    if passthrough_args and isinstance(passthrough_args, six.text_type):
      passthrough_args = passthrough_args.split(',')
    # With --non-telemetry, the gtest executable file path will be passed in as
    # options.executable, which is different from running on shard map. Thus,
    # we don't override executable as we do in running on shard map.
    command_generator = GtestCommandGenerator(
        options, additional_flags=passthrough_args, ignore_shard_env_vars=True)
    # Fallback to use the name of the executable if flag isn't set.
    # TODO(crbug.com/870899): remove fallback logic and raise parser error if
    # --non-telemetry is set but --gtest-benchmark-name is not set once pinpoint
    # is converted to always pass --gtest-benchmark-name flag.
    if not benchmark_name:
      benchmark_name = options.executable
    output_paths = OutputFilePaths(isolated_out_dir, benchmark_name).SetUp()
    print('\n### {folder} ###'.format(folder=benchmark_name))
    overall_return_code = execute_gtest_perf_test(
        command_generator, output_paths, options.xvfb)
    test_results_files.append(output_paths.test_results)
  else:
    if options.use_dynamic_shards:
      shard_map_str = options.dynamic_shardmap
      shard_map = json.loads(shard_map_str, object_pairs_hook=OrderedDict)
      shard_map_path = os.path.join(SHARD_MAPS_DIRECTORY,
                                    options.test_shard_map_filename)
      with open(shard_map_path, 'w') as f:
        json.dump(shard_map, f, indent=4, separators=(',', ': '))
      shutil.copyfile(
          shard_map_path,
          os.path.join(isolated_out_dir, 'benchmarks_shard_map.json'))
      overall_return_code = _run_benchmarks_on_shardmap(
          shard_map, options, isolated_out_dir, test_results_files
      )
    # If the user has supplied a list of benchmark names, execute those instead
    # of using the shard map.
    elif options.benchmarks:
      benchmarks = options.benchmarks.split(',')
      for benchmark in benchmarks:
        output_paths = OutputFilePaths(isolated_out_dir, benchmark).SetUp()
        command_generator = TelemetryCommandGenerator(
            benchmark, options)
        print('\n### {folder} ###'.format(folder=benchmark))
        return_code = execute_telemetry_benchmark(
            command_generator, output_paths, options.xvfb)
        overall_return_code = return_code or overall_return_code
        test_results_files.append(output_paths.test_results)
      if options.run_ref_build:
        print('Not running reference build. --run-ref-build argument is only '
              'supported for sharded benchmarks. It is simple to support '
              'this for unsharded --benchmarks if needed.')
    elif options.test_shard_map_filename:
      # First determine what shard we are running on to know how to
      # index into the bot map to get list of telemetry benchmarks to run.
      shard_map_path = os.path.join(SHARD_MAPS_DIRECTORY,
                                    options.test_shard_map_filename)
      # Copy sharding map file to isolated_out_dir so that the merge script
      # can collect it later.
      shutil.copyfile(
          shard_map_path,
          os.path.join(isolated_out_dir, 'benchmarks_shard_map.json'))
      with open(shard_map_path) as f:
        shard_map = json.load(f)
      overall_return_code = _run_benchmarks_on_shardmap(
          shard_map, options, isolated_out_dir, test_results_files
      )
    else:
      raise Exception('Telemetry tests must provide either a shard map or a '
                      '--benchmarks list so that we know which stories to run.')

  test_results_list = []
  for test_results_file in test_results_files:
    if os.path.exists(test_results_file):
      with open(test_results_file, 'r') as fh:
        test_results_list.append(json.load(fh))
  merged_test_results = results_merger.merge_test_results(test_results_list)
  with open(options.isolated_script_test_output, 'w') as f:
    json.dump(merged_test_results, f)

  return overall_return_code

def _run_benchmarks_on_shardmap(
    shard_map, options, isolated_out_dir, test_results_files):
  overall_return_code = 0
  shard_index = None
  env = os.environ.copy()
  if 'GTEST_SHARD_INDEX' in env:
    shard_index = env['GTEST_SHARD_INDEX']
  # TODO(crbug.com/972844): shard environment variables are not specified
  # for single-shard shard runs.
  if not shard_index:
    shard_map_has_multiple_shards = bool(shard_map.get('1', False))
    if not shard_map_has_multiple_shards:
      shard_index = '0'
  if not shard_index:
    raise Exception(
        'Sharded Telemetry perf tests must either specify --benchmarks '
        'list or have GTEST_SHARD_INDEX environment variable present.')
  shard_configuration = shard_map[shard_index]
  assert ('benchmarks' in shard_configuration or
          'executables' in shard_configuration), (
      'Every shard must have benchmarks or executables associated '
      'with it.')
  if 'benchmarks' in shard_configuration:
    benchmarks_and_configs = shard_configuration['benchmarks']
    for (benchmark, story_selection_config) in benchmarks_and_configs.items():
      # Need to run the benchmark on both latest browser and reference
      # build.
      output_paths = OutputFilePaths(isolated_out_dir, benchmark).SetUp()
      command_generator = TelemetryCommandGenerator(
          benchmark, options,
          story_selection_config=story_selection_config)
      print('\n### {folder} ###'.format(folder=benchmark))
      return_code = execute_telemetry_benchmark(
          command_generator, output_paths, options.xvfb)
      overall_return_code = return_code or overall_return_code
      test_results_files.append(output_paths.test_results)
      if options.run_ref_build:
        reference_benchmark_foldername = benchmark + '.reference'
        reference_output_paths = OutputFilePaths(
            isolated_out_dir, reference_benchmark_foldername).SetUp()
        reference_command_generator = TelemetryCommandGenerator(
            benchmark, options,
            story_selection_config=story_selection_config,
            is_reference=True)
        print('\n### {folder} ###'.format(
            folder=reference_benchmark_foldername))
        # We intentionally ignore the return code and test results of the
        # reference build.
        execute_telemetry_benchmark(
            reference_command_generator, reference_output_paths,
            options.xvfb)
  if 'executables' in shard_configuration:
    names_and_configs = shard_configuration['executables']
    for (name, configuration) in names_and_configs.items():
      additional_flags = []
      if 'arguments' in configuration:
        additional_flags = configuration['arguments']
      command_generator = GtestCommandGenerator(
          options, override_executable=configuration['path'],
          additional_flags=additional_flags, ignore_shard_env_vars=True)
      output_paths = OutputFilePaths(isolated_out_dir, name).SetUp()
      print('\n### {folder} ###'.format(folder=name))
      return_code = execute_gtest_perf_test(
          command_generator, output_paths, options.xvfb)
      overall_return_code = return_code or overall_return_code
      test_results_files.append(output_paths.test_results)

  return overall_return_code

# This is not really a "script test" so does not need to manually add
# any additional compile targets.
def main_compile_targets(args):
  json.dump([], args.output)


if __name__ == '__main__':
  # Conform minimally to the protocol defined by ScriptTest.
  if 'compile_targets' in sys.argv:
    funcs = {
      'run': None,
      'compile_targets': main_compile_targets,
    }
    sys.exit(common.run_script(sys.argv[1:], funcs))

  sys.exit(main(sys.argv))