#!/usr/bin/env vpython # # Copyright 2021 The ANGLE Project Authors. All rights reserved. # Use of this source code is governed by a BSD-style license that can be # found in the LICENSE file. # # process_angle_perf_results.py: # Perf result merging and upload. Adapted from the Chromium script: # https://chromium.googlesource.com/chromium/src/+/main/tools/perf/process_perf_results.py from __future__ import print_function import argparse import collections import datetime import json import logging import multiprocessing import os import pathlib import shutil import subprocess import sys import tempfile import time import uuid logging.basicConfig( level=logging.INFO, format='(%(levelname)s) %(asctime)s pid=%(process)d' ' %(module)s.%(funcName)s:%(lineno)d %(message)s') PY_UTILS = str(pathlib.Path(__file__).resolve().parents[1] / 'src' / 'tests' / 'py_utils') if PY_UTILS not in sys.path: os.stat(PY_UTILS) and sys.path.insert(0, PY_UTILS) import angle_metrics import angle_path_util angle_path_util.AddDepsDirToPath('tools/perf') from core import path_util path_util.AddTelemetryToPath() from core import upload_results_to_perf_dashboard from core import results_merger path_util.AddAndroidPylibToPath() try: from pylib.utils import logdog_helper except ImportError: pass path_util.AddTracingToPath() from tracing.value import histogram from tracing.value import histogram_set from tracing.value.diagnostics import generic_set from tracing.value.diagnostics import reserved_infos RESULTS_URL = 'https://chromeperf.appspot.com' JSON_CONTENT_TYPE = 'application/json' MACHINE_GROUP = 'ANGLE' BUILD_URL = 'https://ci.chromium.org/ui/p/angle/builders/ci/%s/%d' GSUTIL_PY_PATH = str( pathlib.Path(__file__).resolve().parents[1] / 'third_party' / 'depot_tools' / 'gsutil.py') def _upload_perf_results(json_to_upload, name, configuration_name, build_properties, output_json_file): """Upload the contents of result JSON(s) to the perf dashboard.""" args = [ '--buildername', build_properties['buildername'], '--buildnumber', str(build_properties['buildnumber']), '--name', name, '--configuration-name', configuration_name, '--results-file', json_to_upload, '--results-url', RESULTS_URL, '--output-json-file', output_json_file, '--perf-dashboard-machine-group', MACHINE_GROUP, '--got-angle-revision', build_properties['got_angle_revision'], '--send-as-histograms', '--project', 'angle', ] if build_properties.get('git_revision'): args.append('--git-revision') args.append(build_properties['git_revision']) #TODO(crbug.com/1072729): log this in top level logging.info('upload_results_to_perf_dashboard: %s.' % args) return upload_results_to_perf_dashboard.main(args) def _merge_json_output(output_json, jsons_to_merge, extra_links, test_cross_device=False): """Merges the contents of one or more results JSONs. Args: output_json: A path to a JSON file to which the merged results should be written. jsons_to_merge: A list of JSON files that should be merged. extra_links: a (key, value) map in which keys are the human-readable strings which describe the data, and value is logdog url that contain the data. """ begin_time = time.time() merged_results = results_merger.merge_test_results(jsons_to_merge, test_cross_device) # Only append the perf results links if present if extra_links: merged_results['links'] = extra_links with open(output_json, 'w') as f: json.dump(merged_results, f) end_time = time.time() print_duration('Merging json test results', begin_time, end_time) return 0 def _handle_perf_json_test_results(benchmark_directory_map, test_results_list): """Checks the test_results.json under each folder: 1. mark the benchmark 'enabled' if tests results are found 2. add the json content to a list for non-ref. """ begin_time = time.time() benchmark_enabled_map = {} for benchmark_name, directories in benchmark_directory_map.items(): for directory in directories: # Obtain the test name we are running is_ref = '.reference' in benchmark_name enabled = True try: with open(os.path.join(directory, 'test_results.json')) as json_data: json_results = json.load(json_data) if not json_results: # Output is null meaning the test didn't produce any results. # Want to output an error and continue loading the rest of the # test results. logging.warning('No results produced for %s, skipping upload' % directory) continue if json_results.get('version') == 3: # Non-telemetry tests don't have written json results but # if they are executing then they are enabled and will generate # chartjson results. if not bool(json_results.get('tests')): enabled = False if not is_ref: # We don't need to upload reference build data to the # flakiness dashboard since we don't monitor the ref build test_results_list.append(json_results) except IOError as e: # TODO(crbug.com/936602): Figure out how to surface these errors. Should # we have a non-zero exit code if we error out? logging.error('Failed to obtain test results for %s: %s', benchmark_name, e) continue if not enabled: # We don't upload disabled benchmarks or tests that are run # as a smoke test logging.info('Benchmark %s ran no tests on at least one shard' % benchmark_name) continue benchmark_enabled_map[benchmark_name] = True end_time = time.time() print_duration('Analyzing perf json test results', begin_time, end_time) return benchmark_enabled_map def _generate_unique_logdog_filename(name_prefix): return name_prefix + '_' + str(uuid.uuid4()) def _handle_perf_logs(benchmark_directory_map, extra_links): """ Upload benchmark logs to logdog and add a page entry for them. """ begin_time = time.time() benchmark_logs_links = collections.defaultdict(list) for benchmark_name, directories in benchmark_directory_map.items(): for directory in directories: benchmark_log_file = os.path.join(directory, 'benchmark_log.txt') if os.path.exists(benchmark_log_file): with open(benchmark_log_file) as f: uploaded_link = logdog_helper.text( name=_generate_unique_logdog_filename(benchmark_name), data=f.read()) benchmark_logs_links[benchmark_name].append(uploaded_link) logdog_file_name = _generate_unique_logdog_filename('Benchmarks_Logs') logdog_stream = logdog_helper.text( logdog_file_name, json.dumps(benchmark_logs_links, sort_keys=True, indent=4, separators=(',', ': ')), content_type=JSON_CONTENT_TYPE) extra_links['Benchmarks logs'] = logdog_stream end_time = time.time() print_duration('Generating perf log streams', begin_time, end_time) def _handle_benchmarks_shard_map(benchmarks_shard_map_file, extra_links): begin_time = time.time() with open(benchmarks_shard_map_file) as f: benchmarks_shard_data = f.read() logdog_file_name = _generate_unique_logdog_filename('Benchmarks_Shard_Map') logdog_stream = logdog_helper.text( logdog_file_name, benchmarks_shard_data, content_type=JSON_CONTENT_TYPE) extra_links['Benchmarks shard map'] = logdog_stream end_time = time.time() print_duration('Generating benchmark shard map stream', begin_time, end_time) def _get_benchmark_name(directory): return os.path.basename(directory).replace(" benchmark", "") def _scan_output_dir(task_output_dir): benchmark_directory_map = {} benchmarks_shard_map_file = None directory_list = [ f for f in os.listdir(task_output_dir) if not os.path.isfile(os.path.join(task_output_dir, f)) ] benchmark_directory_list = [] for directory in directory_list: for f in os.listdir(os.path.join(task_output_dir, directory)): path = os.path.join(task_output_dir, directory, f) if os.path.isdir(path): benchmark_directory_list.append(path) elif path.endswith('benchmarks_shard_map.json'): benchmarks_shard_map_file = path # Now create a map of benchmark name to the list of directories # the lists were written to. for directory in benchmark_directory_list: benchmark_name = _get_benchmark_name(directory) logging.debug('Found benchmark %s directory %s' % (benchmark_name, directory)) if benchmark_name in benchmark_directory_map.keys(): benchmark_directory_map[benchmark_name].append(directory) else: benchmark_directory_map[benchmark_name] = [directory] return benchmark_directory_map, benchmarks_shard_map_file def _upload_to_skia_perf(benchmark_directory_map, benchmark_enabled_map, build_properties_map): metric_filenames = [] for benchmark_name, directories in benchmark_directory_map.items(): if not benchmark_enabled_map.get(benchmark_name, False): continue for directory in directories: metric_filenames.append(os.path.join(directory, 'angle_metrics.json')) assert metric_filenames buildername = build_properties_map['buildername'] # e.g. win10-nvidia-gtx1660-perf skia_data = { 'version': 1, 'git_hash': build_properties_map['got_angle_revision'], 'key': { 'buildername': buildername, }, 'results': angle_metrics.ConvertToSkiaPerf(metric_filenames), } skia_perf_dir = tempfile.mkdtemp('skia_perf') try: local_file = os.path.join(skia_perf_dir, '%s.%s.json' % (buildername, time.time())) with open(local_file, 'w') as f: json.dump(skia_data, f, indent=2) gs_dir = 'gs://angle-perf-skia/angle_perftests/%s/' % ( datetime.datetime.now().strftime('%Y/%m/%d/%H')) upload_cmd = ['vpython3', GSUTIL_PY_PATH, 'cp', local_file, gs_dir] logging.info('Skia upload: %s', ' '.join(upload_cmd)) subprocess.check_call(upload_cmd) finally: shutil.rmtree(skia_perf_dir) def process_perf_results(output_json, configuration_name, build_properties, task_output_dir, smoke_test_mode, output_results_dir, lightweight=False, skip_perf=False): """Process perf results. Consists of merging the json-test-format output, uploading the perf test output (histogram), and store the benchmark logs in logdog. Each directory in the task_output_dir represents one benchmark that was run. Within this directory, there is a subdirectory with the name of the benchmark that was run. In that subdirectory, there is a perftest-output.json file containing the performance results in histogram format and an output.json file containing the json test results for the benchmark. Returns: (return_code, upload_results_map): return_code is 0 if the whole operation is successful, non zero otherwise. benchmark_upload_result_map: the dictionary that describe which benchmarks were successfully uploaded. """ handle_perf = not lightweight or not skip_perf handle_non_perf = not lightweight or skip_perf logging.info('lightweight mode: %r; handle_perf: %r; handle_non_perf: %r' % (lightweight, handle_perf, handle_non_perf)) begin_time = time.time() return_code = 0 benchmark_upload_result_map = {} benchmark_directory_map, benchmarks_shard_map_file = _scan_output_dir(task_output_dir) test_results_list = [] extra_links = {} if handle_non_perf: # First, upload benchmarks shard map to logdog and add a page # entry for it in extra_links. if benchmarks_shard_map_file: _handle_benchmarks_shard_map(benchmarks_shard_map_file, extra_links) # Second, upload all the benchmark logs to logdog and add a page entry for # those links in extra_links. _handle_perf_logs(benchmark_directory_map, extra_links) # Then try to obtain the list of json test results to merge # and determine the status of each benchmark. benchmark_enabled_map = _handle_perf_json_test_results(benchmark_directory_map, test_results_list) if not smoke_test_mode and handle_perf: build_properties_map = json.loads(build_properties) if not configuration_name: # we are deprecating perf-id crbug.com/817823 configuration_name = build_properties_map['buildername'] try: return_code, benchmark_upload_result_map = _handle_perf_results( benchmark_enabled_map, benchmark_directory_map, configuration_name, build_properties_map, extra_links, output_results_dir) except Exception: logging.exception('Error handling perf results jsons') return_code = 1 try: _upload_to_skia_perf(benchmark_directory_map, benchmark_enabled_map, build_properties_map) except Exception: logging.exception('Error uploading to skia perf') return_code = 1 if handle_non_perf: # Finally, merge all test results json, add the extra links and write out to # output location try: _merge_json_output(output_json, test_results_list, extra_links) except Exception: logging.exception('Error handling test results jsons.') end_time = time.time() print_duration('Total process_perf_results', begin_time, end_time) return return_code, benchmark_upload_result_map def _merge_histogram_results(histogram_lists): merged_results = [] for histogram_list in histogram_lists: merged_results += histogram_list return merged_results def _load_histogram_set_from_dict(data): histograms = histogram_set.HistogramSet() histograms.ImportDicts(data) return histograms def _add_build_info(results, benchmark_name, build_properties): histograms = _load_histogram_set_from_dict(results) common_diagnostics = { reserved_infos.MASTERS: build_properties['builder_group'], reserved_infos.BOTS: build_properties['buildername'], reserved_infos.POINT_ID: build_properties['angle_commit_pos'], reserved_infos.BENCHMARKS: benchmark_name, reserved_infos.ANGLE_REVISIONS: build_properties['got_angle_revision'], reserved_infos.BUILD_URLS: BUILD_URL % (build_properties['buildername'], build_properties['buildnumber']), } for k, v in common_diagnostics.items(): histograms.AddSharedDiagnosticToAllHistograms(k.name, generic_set.GenericSet([v])) return histograms.AsDicts() def _merge_perf_results(benchmark_name, results_filename, directories, build_properties): begin_time = time.time() collected_results = [] for directory in directories: filename = os.path.join(directory, 'perf_results.json') try: with open(filename) as pf: collected_results.append(json.load(pf)) except IOError as e: # TODO(crbug.com/936602): Figure out how to surface these errors. Should # we have a non-zero exit code if we error out? logging.error('Failed to obtain perf results from %s: %s', directory, e) if not collected_results: logging.error('Failed to obtain any perf results from %s.', benchmark_name) return # Assuming that multiple shards will be histogram set # Non-telemetry benchmarks only ever run on one shard merged_results = [] assert (isinstance(collected_results[0], list)) merged_results = _merge_histogram_results(collected_results) # Write additional histogram build info. merged_results = _add_build_info(merged_results, benchmark_name, build_properties) with open(results_filename, 'w') as rf: json.dump(merged_results, rf) end_time = time.time() print_duration(('%s results merging' % (benchmark_name)), begin_time, end_time) def _upload_individual(benchmark_name, directories, configuration_name, build_properties, output_json_file): tmpfile_dir = tempfile.mkdtemp() try: upload_begin_time = time.time() # There are potentially multiple directores with results, re-write and # merge them if necessary results_filename = None if len(directories) > 1: merge_perf_dir = os.path.join(os.path.abspath(tmpfile_dir), benchmark_name) if not os.path.exists(merge_perf_dir): os.makedirs(merge_perf_dir) results_filename = os.path.join(merge_perf_dir, 'merged_perf_results.json') _merge_perf_results(benchmark_name, results_filename, directories, build_properties) else: # It was only written to one shard, use that shards data results_filename = os.path.join(directories[0], 'perf_results.json') results_size_in_mib = os.path.getsize(results_filename) / (2**20) logging.info('Uploading perf results from %s benchmark (size %s Mib)' % (benchmark_name, results_size_in_mib)) upload_return_code = _upload_perf_results(results_filename, benchmark_name, configuration_name, build_properties, output_json_file) upload_end_time = time.time() print_duration(('%s upload time' % (benchmark_name)), upload_begin_time, upload_end_time) return (benchmark_name, upload_return_code == 0) finally: shutil.rmtree(tmpfile_dir) def _upload_individual_benchmark(params): try: return _upload_individual(*params) except Exception: benchmark_name = params[0] upload_succeed = False logging.exception('Error uploading perf result of %s' % benchmark_name) return benchmark_name, upload_succeed def _GetCpuCount(log=True): try: cpu_count = multiprocessing.cpu_count() if sys.platform == 'win32': # TODO(crbug.com/1190269) - we can't use more than 56 # cores on Windows or Python3 may hang. cpu_count = min(cpu_count, 56) return cpu_count except NotImplementedError: if log: logging.warn('Failed to get a CPU count for this bot. See crbug.com/947035.') # TODO(crbug.com/948281): This is currently set to 4 since the mac masters # only have 4 cores. Once we move to all-linux, this can be increased or # we can even delete this whole function and use multiprocessing.cpu_count() # directly. return 4 def _load_shard_id_from_test_results(directory): shard_id = None test_json_path = os.path.join(directory, 'test_results.json') try: with open(test_json_path) as f: test_json = json.load(f) all_results = test_json['tests'] for _, benchmark_results in all_results.items(): for _, measurement_result in benchmark_results.items(): shard_id = measurement_result['shard'] break except IOError as e: logging.error('Failed to open test_results.json from %s: %s', test_json_path, e) except KeyError as e: logging.error('Failed to locate results in test_results.json: %s', e) return shard_id def _find_device_id_by_shard_id(benchmarks_shard_map_file, shard_id): try: with open(benchmarks_shard_map_file) as f: shard_map_json = json.load(f) device_id = shard_map_json['extra_infos']['bot #%s' % shard_id] except KeyError as e: logging.error('Failed to locate device name in shard map: %s', e) return device_id def _update_perf_json_with_summary_on_device_id(directory, device_id): perf_json_path = os.path.join(directory, 'perf_results.json') try: with open(perf_json_path, 'r') as f: perf_json = json.load(f) except IOError as e: logging.error('Failed to open perf_results.json from %s: %s', perf_json_path, e) summary_key_guid = str(uuid.uuid4()) summary_key_generic_set = { 'values': ['device_id'], 'guid': summary_key_guid, 'type': 'GenericSet' } perf_json.insert(0, summary_key_generic_set) logging.info('Inserted summary key generic set for perf result in %s: %s', directory, summary_key_generic_set) stories_guids = set() for entry in perf_json: if 'diagnostics' in entry: entry['diagnostics']['summaryKeys'] = summary_key_guid stories_guids.add(entry['diagnostics']['stories']) for entry in perf_json: if 'guid' in entry and entry['guid'] in stories_guids: entry['values'].append(device_id) try: with open(perf_json_path, 'w') as f: json.dump(perf_json, f) except IOError as e: logging.error('Failed to writing perf_results.json to %s: %s', perf_json_path, e) logging.info('Finished adding device id %s in perf result.', device_id) def _handle_perf_results(benchmark_enabled_map, benchmark_directory_map, configuration_name, build_properties, extra_links, output_results_dir): """ Upload perf results to the perf dashboard. This method also upload the perf results to logdog and augment it to |extra_links|. Returns: (return_code, benchmark_upload_result_map) return_code is 0 if this upload to perf dashboard successfully, 1 otherwise. benchmark_upload_result_map is a dictionary describes which benchmark was successfully uploaded. """ begin_time = time.time() # Upload all eligible benchmarks to the perf dashboard results_dict = {} invocations = [] for benchmark_name, directories in benchmark_directory_map.items(): if not benchmark_enabled_map.get(benchmark_name, False): continue # Create a place to write the perf results that you will write out to # logdog. output_json_file = os.path.join(output_results_dir, (str(uuid.uuid4()) + benchmark_name)) results_dict[benchmark_name] = output_json_file #TODO(crbug.com/1072729): pass final arguments instead of build properties # and configuration_name invocations.append( (benchmark_name, directories, configuration_name, build_properties, output_json_file)) # Kick off the uploads in multiple processes # crbug.com/1035930: We are hitting HTTP Response 429. Limit ourselves # to 2 processes to avoid this error. Uncomment the following code once # the problem is fixed on the dashboard side. # pool = multiprocessing.Pool(_GetCpuCount()) pool = multiprocessing.Pool(2) upload_result_timeout = False try: async_result = pool.map_async(_upload_individual_benchmark, invocations) # TODO(crbug.com/947035): What timeout is reasonable? results = async_result.get(timeout=4000) except multiprocessing.TimeoutError: upload_result_timeout = True logging.error('Timeout uploading benchmarks to perf dashboard in parallel') results = [] for benchmark_name in benchmark_directory_map: results.append((benchmark_name, False)) finally: pool.terminate() # Keep a mapping of benchmarks to their upload results benchmark_upload_result_map = {} for r in results: benchmark_upload_result_map[r[0]] = r[1] logdog_dict = {} upload_failures_counter = 0 logdog_stream = None logdog_label = 'Results Dashboard' for benchmark_name, output_file in results_dict.items(): upload_succeed = benchmark_upload_result_map[benchmark_name] if not upload_succeed: upload_failures_counter += 1 is_reference = '.reference' in benchmark_name _write_perf_data_to_logfile( benchmark_name, output_file, configuration_name, build_properties, logdog_dict, is_reference, upload_failure=not upload_succeed) logdog_file_name = _generate_unique_logdog_filename('Results_Dashboard_') logdog_stream = logdog_helper.text( logdog_file_name, json.dumps(dict(logdog_dict), sort_keys=True, indent=4, separators=(',', ': ')), content_type=JSON_CONTENT_TYPE) if upload_failures_counter > 0: logdog_label += (' %s merge script perf data upload failures' % upload_failures_counter) extra_links[logdog_label] = logdog_stream end_time = time.time() print_duration('Uploading results to perf dashboard', begin_time, end_time) if upload_result_timeout or upload_failures_counter > 0: return 1, benchmark_upload_result_map return 0, benchmark_upload_result_map def _write_perf_data_to_logfile(benchmark_name, output_file, configuration_name, build_properties, logdog_dict, is_ref, upload_failure): viewer_url = None # logdog file to write perf results to if os.path.exists(output_file): results = None with open(output_file) as f: try: results = json.load(f) except ValueError: logging.error('Error parsing perf results JSON for benchmark %s' % benchmark_name) if results: try: json_fname = _generate_unique_logdog_filename(benchmark_name) output_json_file = logdog_helper.open_text(json_fname) json.dump(results, output_json_file, indent=4, separators=(',', ': ')) except ValueError as e: logging.error('ValueError: "%s" while dumping output to logdog' % e) finally: output_json_file.close() viewer_url = output_json_file.get_viewer_url() else: logging.warning("Perf results JSON file doesn't exist for benchmark %s" % benchmark_name) base_benchmark_name = benchmark_name.replace('.reference', '') if base_benchmark_name not in logdog_dict: logdog_dict[base_benchmark_name] = {} # add links for the perf results and the dashboard url to # the logs section of buildbot if is_ref: if viewer_url: logdog_dict[base_benchmark_name]['perf_results_ref'] = viewer_url if upload_failure: logdog_dict[base_benchmark_name]['ref_upload_failed'] = 'True' else: # TODO(jmadill): Figure out if we can get a dashboard URL here. http://anglebug.com/40096778 # logdog_dict[base_benchmark_name]['dashboard_url'] = ( # upload_results_to_perf_dashboard.GetDashboardUrl(benchmark_name, configuration_name, # RESULTS_URL, # build_properties['got_revision_cp'], # _GetMachineGroup(build_properties))) if viewer_url: logdog_dict[base_benchmark_name]['perf_results'] = viewer_url if upload_failure: logdog_dict[base_benchmark_name]['upload_failed'] = 'True' def print_duration(step, start, end): logging.info('Duration of %s: %d seconds' % (step, end - start)) def main(): """ See collect_task.collect_task for more on the merge script API. """ logging.info(sys.argv) parser = argparse.ArgumentParser() # configuration-name (previously perf-id) is the name of bot the tests run on # For example, buildbot-test is the name of the android-go-perf bot # configuration-name and results-url are set in the json file which is going # away tools/perf/core/chromium.perf.fyi.extras.json parser.add_argument('--configuration-name', help=argparse.SUPPRESS) parser.add_argument('--build-properties', help=argparse.SUPPRESS) parser.add_argument('--summary-json', required=True, help=argparse.SUPPRESS) parser.add_argument('--task-output-dir', required=True, help=argparse.SUPPRESS) parser.add_argument('-o', '--output-json', required=True, help=argparse.SUPPRESS) parser.add_argument( '--skip-perf', action='store_true', help='In lightweight mode, using --skip-perf will skip the performance' ' data handling.') parser.add_argument( '--lightweight', action='store_true', help='Choose the lightweight mode in which the perf result handling' ' is performed on a separate VM.') parser.add_argument('json_files', nargs='*', help=argparse.SUPPRESS) parser.add_argument( '--smoke-test-mode', action='store_true', help='This test should be run in smoke test mode' ' meaning it does not upload to the perf dashboard') args = parser.parse_args() with open(args.summary_json) as f: shard_summary = json.load(f) shard_failed = any(int(shard.get('exit_code', 1)) != 0 for shard in shard_summary['shards']) output_results_dir = tempfile.mkdtemp('outputresults') try: return_code, _ = process_perf_results(args.output_json, args.configuration_name, args.build_properties, args.task_output_dir, args.smoke_test_mode, output_results_dir, args.lightweight, args.skip_perf) except Exception: logging.exception('process_perf_results raised an exception') return_code = 1 finally: shutil.rmtree(output_results_dir) if return_code != 0 and shard_failed: logging.warning('Perf processing failed but one or more shards failed earlier') return_code = 0 # Enables the failed build info to be rendered normally return return_code if __name__ == '__main__': sys.exit(main())