Exemple #1
0
    def _run(self,
             discovery_schedule=None,
             exit_schedule=None,
             exit_mode='exception',
             np=2,
             min_np=2,
             max_np=4,
             hosts=None):
        if not discovery_schedule and not hosts:
            raise ValueError(
                'at least one of discovery schedule or hosts must be given')

        with temppath() as logfile:
            with _temp_discovery_script(logfile, discovery_schedule or [(None, hosts.split(','))]) \
                    as discovery_script:
                command_args = [
                    'horovodrun', '-np',
                    str(np), '--min-np',
                    str(min_np), '--log-level', 'DEBUG'
                ]
                if hosts is not None:
                    command_args += ['-H', hosts]
                else:
                    command_args += [
                        '--host-discovery-script', discovery_script,
                        '--max-np',
                        str(max_np)
                    ]

                command_args += [
                    'python', self._training_script, '--logfile', logfile
                ]
                if discovery_schedule:
                    command_args += [
                        '--discovery-schedule',
                        json.dumps(discovery_schedule)
                    ]
                if exit_schedule:
                    command_args += [
                        '--exit-schedule',
                        json.dumps(exit_schedule), '--exit-mode', exit_mode
                    ]
                print(' '.join(command_args))

                with override_args(*command_args):
                    args = parse_args()
                    env = {}
                    config_parser.set_env_from_args(env, args)
                    _run_elastic(args)

                    with open(logfile, 'r') as f:
                        lines = f.readlines()

                    print('logfile:')
                    for line in lines:
                        print(line)

                    return [json.loads(line) for line in lines]
Exemple #2
0
    def test_logging_args(self):
        with override_args('horovodrun', '-np', '2', '--log-level', 'INFO',
                           '--log-hide-timestamp'):
            args = parse_args()
            env = {}
            config_parser.set_env_from_args(env, args)

            self.assertEqual(env.get(config_parser.HOROVOD_LOG_LEVEL), 'INFO')
            self.assertEqual(env.get(config_parser.HOROVOD_LOG_HIDE_TIME), '1')
Exemple #3
0
 def test_config_file_override_args(self):
     config_filename = os.path.join(os.path.dirname(__file__), 'data/config.test.yaml')
     with override_args('horovodrun', '-np', '2',
                        '--fusion-threshold-mb', '128',
                        '--config-file', config_filename,
                        '--cycle-time-ms', '20',):
         args = parse_args()
         self.assertEqual(args.fusion_threshold_mb, 128)
         self.assertEqual(args.cycle_time_ms, 20)
Exemple #4
0
    def test_stall_check_args(self):
        with override_args('horovodrun', '-np', '2',
                           '--no-stall-check'):
            args = parse_args()
            env = {}
            config_parser.set_env_from_args(env, args)

            self.assertEqual(env.get(config_parser.HOROVOD_STALL_CHECK_DISABLE), '1')

        with override_args('horovodrun', '-np', '2',
                           '--stall-check-warning-time-seconds', '10',
                           '--stall-check-shutdown-time-seconds', '20'):
            args = parse_args()
            env = {}
            config_parser.set_env_from_args(env, args)

            self.assertNotIn(config_parser.HOROVOD_STALL_CHECK_DISABLE, env)
            self.assertEqual(env.get(config_parser.HOROVOD_STALL_CHECK_TIME_SECONDS), '10')
            self.assertEqual(env.get(config_parser.HOROVOD_STALL_SHUTDOWN_TIME_SECONDS), '20')
Exemple #5
0
    def test_timeline_args(self):
        with override_args('horovodrun', '-np', '2',
                           '--timeline-filename', '/tmp/timeline.json',
                           '--timeline-mark-cycles'):
            args = parse_args()
            env = {}
            config_parser.set_env_from_args(env, args)

            self.assertEqual(env.get(config_parser.HOROVOD_TIMELINE), '/tmp/timeline.json')
            self.assertEqual(env.get(config_parser.HOROVOD_TIMELINE_MARK_CYCLES), '1')
Exemple #6
0
    def test_autotuning_with_fixed_param(self):
        with override_args('horovodrun', '-np', '2',
                           '--autotune',
                           '--cache-capacity', '1024',
                           '--no-hierarchical-allgather'):
            args = parse_args()
            env = {}
            config_parser.set_env_from_args(env, args)

            self.assertNotIn(config_parser.HOROVOD_FUSION_THRESHOLD, env)
            self.assertNotIn(config_parser.HOROVOD_CYCLE_TIME, env)
            self.assertEqual(env.get(config_parser.HOROVOD_CACHE_CAPACITY), '1024')
            self.assertNotIn(config_parser.HOROVOD_HIERARCHICAL_ALLREDUCE, env)
            self.assertEqual(env.get(config_parser.HOROVOD_HIERARCHICAL_ALLGATHER), '0')
Exemple #7
0
    def test_library_args(self):
        with override_args('horovodrun', '-np', '2',
                           '--mpi-threads-disable',
                           '--num-nccl-streams', '2',
                           '--ccl-bgt-affinity', '1',
                           '--gloo-timeout-seconds', '60'):
            args = parse_args()
            env = {}
            config_parser.set_env_from_args(env, args)

            self.assertEqual(env.get(config_parser.HOROVOD_MPI_THREADS_DISABLE), '1')
            self.assertEqual(env.get(config_parser.HOROVOD_NUM_NCCL_STREAMS), '2')
            self.assertEqual(env.get(config_parser.HOROVOD_CCL_BGT_AFFINITY), '1')
            self.assertEqual(env.get(config_parser.HOROVOD_GLOO_TIMEOUT_SECONDS), '60')
Exemple #8
0
    def test_params_args(self):
        with override_args('horovodrun', '-np', '2',
                           '--fusion-threshold-mb', '10',
                           '--cycle-time-ms', '20',
                           '--cache-capacity', '512',
                           '--hierarchical-allreduce',
                           '--hierarchical-allgather'):
            args = parse_args()
            env = {}
            config_parser.set_env_from_args(env, args)

            self.assertEqual(env.get(config_parser.HOROVOD_FUSION_THRESHOLD), str(10 * 1024 * 1024))
            self.assertEqual(env.get(config_parser.HOROVOD_CYCLE_TIME), '20.0')
            self.assertEqual(env.get(config_parser.HOROVOD_CACHE_CAPACITY), '512')
            self.assertEqual(env.get(config_parser.HOROVOD_HIERARCHICAL_ALLREDUCE), '1')
            self.assertEqual(env.get(config_parser.HOROVOD_HIERARCHICAL_ALLGATHER), '1')
Exemple #9
0
    def test_autotune_args(self):
        with override_args('horovodrun', '-np', '2',
                           '--autotune',
                           '--autotune-log-file', '/tmp/autotune.txt',
                           '--autotune-warmup-samples', '1',
                           '--autotune-steps-per-sample', '5',
                           '--autotune-bayes-opt-max-samples', '10',
                           '--autotune-gaussian-process-noise', '0.2'):
            args = parse_args()
            env = {}
            config_parser.set_env_from_args(env, args)

            self.assertEqual(env.get(config_parser.HOROVOD_AUTOTUNE), '1')
            self.assertEqual(env.get(config_parser.HOROVOD_AUTOTUNE_LOG), '/tmp/autotune.txt')
            self.assertEqual(env.get(config_parser.HOROVOD_AUTOTUNE_WARMUP_SAMPLES), '1')
            self.assertEqual(env.get(config_parser.HOROVOD_AUTOTUNE_STEPS_PER_SAMPLE), '5')
            self.assertEqual(env.get(config_parser.HOROVOD_AUTOTUNE_BAYES_OPT_MAX_SAMPLES), '10')
            self.assertEqual(env.get(config_parser.HOROVOD_AUTOTUNE_GAUSSIAN_PROCESS_NOISE), '0.2')
Exemple #10
0
    def test_config_file(self):
        config_filename = os.path.join(os.path.dirname(__file__),
                                       'data/config.test.yaml')
        with override_args('horovodrun', '-np', '2', '--config-file',
                           config_filename):
            args = parse_args()

            self.assertTrue(args.use_gloo)

            # Params
            self.assertEqual(args.fusion_threshold_mb, 32)
            self.assertEqual(args.cycle_time_ms, 10)
            self.assertEqual(args.cache_capacity, 2048)
            self.assertTrue(args.hierarchical_allreduce)
            self.assertTrue(args.hierarchical_allgather)

            # Autotune
            self.assertTrue(args.autotune)
            self.assertEqual(args.autotune_log_file,
                             'horovod_autotune_log.txt')
            self.assertEqual(args.autotune_warmup_samples, 5)
            self.assertEqual(args.autotune_steps_per_sample, 20)
            self.assertEqual(args.autotune_bayes_opt_max_samples, 50)
            self.assertEqual(args.autotune_gaussian_process_noise, 0.9)

            # Timeline
            self.assertEqual(args.timeline_filename, 'horovod_timeline.json')
            self.assertTrue(args.timeline_mark_cycles)

            # Stall Check
            self.assertFalse(args.no_stall_check)
            self.assertEqual(args.stall_check_warning_time_seconds, 120)
            self.assertEqual(args.stall_check_shutdown_time_seconds, 240)

            # Library Options
            self.assertTrue(args.mpi_threads_disable)
            self.assertEqual(args.num_nccl_streams, 2)
            self.assertEqual(args.ccl_bgt_affinity, 1)
            self.assertEqual(args.gloo_timeout_seconds, 60)

            # Logging
            self.assertEqual(args.log_level, 'INFO')
            self.assertTrue(args.log_hide_timestamp)
Exemple #11
0
 def test_validate_config_args(self):
     with override_args('horovodrun', '-np', '2', '--fusion-threshold-mb',
                        '-1'):
         with pytest.raises(ValueError):
             parse_args()