1#!/usr/bin/env python 2# Copyright 2017 gRPC authors. 3# 4# Licensed under the Apache License, Version 2.0 (the "License"); 5# you may not use this file except in compliance with the License. 6# You may obtain a copy of the License at 7# 8# http://www.apache.org/licenses/LICENSE-2.0 9# 10# Unless required by applicable law or agreed to in writing, software 11# distributed under the License is distributed on an "AS IS" BASIS, 12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13# See the License for the specific language governing permissions and 14# limitations under the License. 15 16import cgi 17import multiprocessing 18import os 19import subprocess 20import sys 21import argparse 22 23import python_utils.jobset as jobset 24import python_utils.start_port_server as start_port_server 25 26sys.path.append( 27 os.path.join(os.path.dirname(sys.argv[0]), '..', 'profiling', 28 'microbenchmarks', 'bm_diff')) 29import bm_constants 30 31flamegraph_dir = os.path.join(os.path.expanduser('~'), 'FlameGraph') 32 33os.chdir(os.path.join(os.path.dirname(sys.argv[0]), '../..')) 34if not os.path.exists('reports'): 35 os.makedirs('reports') 36 37start_port_server.start_port_server() 38 39 40def fnize(s): 41 out = '' 42 for c in s: 43 if c in '<>, /': 44 if len(out) and out[-1] == '_': continue 45 out += '_' 46 else: 47 out += c 48 return out 49 50 51# index html 52index_html = """ 53<html> 54<head> 55<title>Microbenchmark Results</title> 56</head> 57<body> 58""" 59 60 61def heading(name): 62 global index_html 63 index_html += "<h1>%s</h1>\n" % name 64 65 66def link(txt, tgt): 67 global index_html 68 index_html += "<p><a href=\"%s\">%s</a></p>\n" % (cgi.escape( 69 tgt, quote=True), cgi.escape(txt)) 70 71 72def text(txt): 73 global index_html 74 index_html += "<p><pre>%s</pre></p>\n" % cgi.escape(txt) 75 76 77def collect_latency(bm_name, args): 78 """generate latency profiles""" 79 benchmarks = [] 80 profile_analysis = [] 81 cleanup = [] 82 83 heading('Latency Profiles: %s' % bm_name) 84 subprocess.check_call([ 85 'make', bm_name, 'CONFIG=basicprof', '-j', 86 '%d' % multiprocessing.cpu_count() 87 ]) 88 for line in subprocess.check_output( 89 ['bins/basicprof/%s' % bm_name, '--benchmark_list_tests']).splitlines(): 90 link(line, '%s.txt' % fnize(line)) 91 benchmarks.append( 92 jobset.JobSpec([ 93 'bins/basicprof/%s' % bm_name, 94 '--benchmark_filter=^%s$' % line, '--benchmark_min_time=0.05' 95 ], 96 environ={ 97 'GRPC_LATENCY_TRACE': '%s.trace' % fnize(line) 98 }, 99 shortname='profile-%s' % fnize(line))) 100 profile_analysis.append( 101 jobset.JobSpec([ 102 sys.executable, 103 'tools/profiling/latency_profile/profile_analyzer.py', 104 '--source', 105 '%s.trace' % fnize(line), '--fmt', 'simple', '--out', 106 'reports/%s.txt' % fnize(line) 107 ], 108 timeout_seconds=20 * 60, 109 shortname='analyze-%s' % fnize(line))) 110 cleanup.append(jobset.JobSpec(['rm', '%s.trace' % fnize(line)])) 111 # periodically flush out the list of jobs: profile_analysis jobs at least 112 # consume upwards of five gigabytes of ram in some cases, and so analysing 113 # hundreds of them at once is impractical -- but we want at least some 114 # concurrency or the work takes too long 115 if len(benchmarks) >= min(16, multiprocessing.cpu_count()): 116 # run up to half the cpu count: each benchmark can use up to two cores 117 # (one for the microbenchmark, one for the data flush) 118 jobset.run(benchmarks, 119 maxjobs=max(1, 120 multiprocessing.cpu_count() / 2)) 121 jobset.run(profile_analysis, maxjobs=multiprocessing.cpu_count()) 122 jobset.run(cleanup, maxjobs=multiprocessing.cpu_count()) 123 benchmarks = [] 124 profile_analysis = [] 125 cleanup = [] 126 # run the remaining benchmarks that weren't flushed 127 if len(benchmarks): 128 jobset.run(benchmarks, maxjobs=max(1, multiprocessing.cpu_count() / 2)) 129 jobset.run(profile_analysis, maxjobs=multiprocessing.cpu_count()) 130 jobset.run(cleanup, maxjobs=multiprocessing.cpu_count()) 131 132 133def collect_perf(bm_name, args): 134 """generate flamegraphs""" 135 heading('Flamegraphs: %s' % bm_name) 136 subprocess.check_call([ 137 'make', bm_name, 'CONFIG=mutrace', '-j', 138 '%d' % multiprocessing.cpu_count() 139 ]) 140 benchmarks = [] 141 profile_analysis = [] 142 cleanup = [] 143 for line in subprocess.check_output( 144 ['bins/mutrace/%s' % bm_name, '--benchmark_list_tests']).splitlines(): 145 link(line, '%s.svg' % fnize(line)) 146 benchmarks.append( 147 jobset.JobSpec([ 148 'perf', 'record', '-o', 149 '%s-perf.data' % fnize(line), '-g', '-F', '997', 150 'bins/mutrace/%s' % bm_name, 151 '--benchmark_filter=^%s$' % line, '--benchmark_min_time=10' 152 ], 153 shortname='perf-%s' % fnize(line))) 154 profile_analysis.append( 155 jobset.JobSpec( 156 [ 157 'tools/run_tests/performance/process_local_perf_flamegraphs.sh' 158 ], 159 environ={ 160 'PERF_BASE_NAME': fnize(line), 161 'OUTPUT_DIR': 'reports', 162 'OUTPUT_FILENAME': fnize(line), 163 }, 164 shortname='flame-%s' % fnize(line))) 165 cleanup.append(jobset.JobSpec(['rm', '%s-perf.data' % fnize(line)])) 166 cleanup.append(jobset.JobSpec(['rm', '%s-out.perf' % fnize(line)])) 167 # periodically flush out the list of jobs: temporary space required for this 168 # processing is large 169 if len(benchmarks) >= 20: 170 # run up to half the cpu count: each benchmark can use up to two cores 171 # (one for the microbenchmark, one for the data flush) 172 jobset.run(benchmarks, maxjobs=1) 173 jobset.run(profile_analysis, maxjobs=multiprocessing.cpu_count()) 174 jobset.run(cleanup, maxjobs=multiprocessing.cpu_count()) 175 benchmarks = [] 176 profile_analysis = [] 177 cleanup = [] 178 # run the remaining benchmarks that weren't flushed 179 if len(benchmarks): 180 jobset.run(benchmarks, maxjobs=1) 181 jobset.run(profile_analysis, maxjobs=multiprocessing.cpu_count()) 182 jobset.run(cleanup, maxjobs=multiprocessing.cpu_count()) 183 184 185def run_summary(bm_name, cfg, base_json_name): 186 subprocess.check_call([ 187 'make', bm_name, 188 'CONFIG=%s' % cfg, '-j', 189 '%d' % multiprocessing.cpu_count() 190 ]) 191 cmd = [ 192 'bins/%s/%s' % (cfg, bm_name), 193 '--benchmark_out=%s.%s.json' % (base_json_name, cfg), 194 '--benchmark_out_format=json' 195 ] 196 if args.summary_time is not None: 197 cmd += ['--benchmark_min_time=%d' % args.summary_time] 198 return subprocess.check_output(cmd) 199 200 201def collect_summary(bm_name, args): 202 heading('Summary: %s [no counters]' % bm_name) 203 text(run_summary(bm_name, 'opt', bm_name)) 204 heading('Summary: %s [with counters]' % bm_name) 205 text(run_summary(bm_name, 'counters', bm_name)) 206 if args.bigquery_upload: 207 with open('%s.csv' % bm_name, 'w') as f: 208 f.write( 209 subprocess.check_output([ 210 'tools/profiling/microbenchmarks/bm2bq.py', 211 '%s.counters.json' % bm_name, 212 '%s.opt.json' % bm_name 213 ])) 214 subprocess.check_call([ 215 'bq', 'load', 'microbenchmarks.microbenchmarks', 216 '%s.csv' % bm_name 217 ]) 218 219 220collectors = { 221 'latency': collect_latency, 222 'perf': collect_perf, 223 'summary': collect_summary, 224} 225 226argp = argparse.ArgumentParser(description='Collect data from microbenchmarks') 227argp.add_argument('-c', 228 '--collect', 229 choices=sorted(collectors.keys()), 230 nargs='*', 231 default=sorted(collectors.keys()), 232 help='Which collectors should be run against each benchmark') 233argp.add_argument('-b', 234 '--benchmarks', 235 choices=bm_constants._AVAILABLE_BENCHMARK_TESTS, 236 default=bm_constants._AVAILABLE_BENCHMARK_TESTS, 237 nargs='+', 238 type=str, 239 help='Which microbenchmarks should be run') 240argp.add_argument('--bigquery_upload', 241 default=False, 242 action='store_const', 243 const=True, 244 help='Upload results from summary collection to bigquery') 245argp.add_argument( 246 '--summary_time', 247 default=None, 248 type=int, 249 help='Minimum time to run benchmarks for the summary collection') 250args = argp.parse_args() 251 252try: 253 for collect in args.collect: 254 for bm_name in args.benchmarks: 255 collectors[collect](bm_name, args) 256finally: 257 if not os.path.exists('reports'): 258 os.makedirs('reports') 259 index_html += "</body>\n</html>\n" 260 with open('reports/index.html', 'w') as f: 261 f.write(index_html) 262