1# Copyright 2018 The TensorFlow Authors. All Rights Reserved. 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# =================================================================== 15"""TPU system metadata and associated tooling.""" 16 17import collections 18 19from tensorflow.core.protobuf import config_pb2 20from tensorflow.python.client import session as session_lib 21from tensorflow.python.distribute import device_util 22from tensorflow.python.eager import context 23from tensorflow.python.framework import config 24from tensorflow.python.framework import device as tf_device 25from tensorflow.python.framework import errors 26from tensorflow.python.framework import ops 27from tensorflow.python.platform import tf_logging as logging 28from tensorflow.python.tpu import tpu 29from tensorflow.python.util.tf_export import tf_export 30 31_PINGING_MASTER_TIMEOUT_IN_MS = 5 * 60 * 1000 # 10 min 32_RETRY_TIMES = 12 * 24 # 1 day 33_INITIAL_TPU_SYSTEM_TIMEOUT_IN_MS = 300 * 1000 # 5 mins 34 35_DEFAULT_JOB_NAME = 'tpu_worker' 36_DEFAULT_COORDINATOR_JOB_NAME = 'coordinator' 37_LOCAL_MASTERS = ('', 'local') 38 39 40@tf_export('tpu.experimental.TPUSystemMetadata') 41class TPUSystemMetadata( 42 collections.namedtuple('TPUSystemMetadata', [ 43 'num_cores', 44 'num_hosts', 45 'num_of_cores_per_host', 46 'topology', 47 'devices', 48 ])): 49 """Describes some metadata about the TPU system. 50 51 Attributes: 52 num_cores: interger. Total number of TPU cores in the TPU system. 53 num_hosts: interger. Total number of hosts (TPU workers) in the TPU system. 54 num_of_cores_per_host: interger. Number of TPU cores per host (TPU worker). 55 topology: an instance of `tf.tpu.experimental.Topology`, which describes the 56 physical topology of TPU system. 57 devices: a tuple of strings, which describes all the TPU devices in the 58 system. 59 """ 60 61 def __new__(cls, num_cores, num_hosts, num_of_cores_per_host, topology, 62 devices): 63 return super(TPUSystemMetadata, 64 cls).__new__(cls, num_cores, num_hosts, num_of_cores_per_host, 65 topology, devices) 66 67 68def _query_tpu_system_metadata(master_address, cluster_def=None, 69 query_topology=False): 70 """Automatically detects the TPU system metadata in the system.""" 71 tpu_core_count = 0 72 devices = [] 73 device_dict = collections.defaultdict(list) 74 75 if context.executing_eagerly(): 76 logical_devices = config.list_logical_devices() 77 78 # We want the output type to match in both eager and session mode 79 devices = [session_lib._DeviceAttributes(device_util.canonicalize(d.name), # pylint: disable=protected-access 80 d.device_type, 0, 0) 81 for d in logical_devices] 82 else: 83 # TODO(b/120564445): Replace with standard library for retries. 84 retry_count = 1 85 while True: 86 logging.info('Querying Tensorflow master (%s) for TPU system metadata.', 87 master_address) 88 try: 89 with ops.Graph().as_default(): 90 with session_lib.Session( 91 master_address, 92 config=get_session_config_with_timeout( 93 _PINGING_MASTER_TIMEOUT_IN_MS, 94 cluster_def)) as sess: 95 devices = sess.list_devices() 96 break 97 except errors.DeadlineExceededError: 98 msg = ('Failed to connect to the Tensorflow master. The TPU worker may ' 99 'not be ready (still scheduling) or the Tensorflow master ' 100 'address is incorrect: got (%s).' % 101 (master_address)) 102 103 # TODO(xiejw): For local or grpc master we might not need retry logic 104 # here. 105 if retry_count <= _RETRY_TIMES: 106 logging.warning('%s', msg) 107 logging.warning('Retrying (%d/%d).', retry_count, _RETRY_TIMES) 108 retry_count += 1 109 else: 110 raise ValueError(msg) 111 112 for device in devices: 113 spec = tf_device.DeviceSpec.from_string(device.name) 114 if spec.device_type == 'TPU': 115 device_dict[spec.task].append(spec.device_index) 116 tpu_core_count += 1 117 118 num_of_cores_per_host = 0 119 if tpu_core_count: 120 num_cores_per_host_set = set( 121 [len(core_ids) for core_ids in device_dict.values()]) 122 if len(num_cores_per_host_set) != 1: 123 raise RuntimeError( 124 'TPU cores on each host is not same. This should not happen!. ' 125 'devices: {}'.format(devices)) 126 num_of_cores_per_host = num_cores_per_host_set.pop() 127 128 topology = None 129 if query_topology: 130 if not tpu_core_count: 131 raise RuntimeError( 132 'Cannot find any TPU cores in the system (master address {}). ' 133 'This usually means the master address is incorrect or the ' 134 'TPU worker has some problems. Available devices: {}'.format( 135 master_address, devices)) 136 137 topology = _obtain_topology(master_address, cluster_def) 138 139 # We sort the metadata devices so that downstream users get a sorted list 140 # for creating mirrored variables correctly. 141 def _sort_key(device): 142 spec = tf_device.DeviceSpec.from_string(device.name) 143 return (spec.job, spec.replica, spec.task, spec.device_type, 144 spec.device_index) 145 devices = tuple(sorted(devices, key=_sort_key)) 146 147 metadata = TPUSystemMetadata( 148 num_cores=tpu_core_count, 149 num_hosts=len(device_dict), 150 num_of_cores_per_host=num_of_cores_per_host, 151 topology=topology, 152 devices=devices) 153 154 if tpu_core_count: 155 logging.info('Found TPU system:') 156 logging.info('*** Num TPU Cores: %d', metadata.num_cores) 157 logging.info('*** Num TPU Workers: %d', metadata.num_hosts) 158 logging.info('*** Num TPU Cores Per Worker: %d', 159 metadata.num_of_cores_per_host) 160 for device in metadata.devices: 161 logging.info('*** Available Device: %s', device) 162 else: 163 logging.info('Failed to find TPU: %s', metadata) 164 return metadata 165 166 167def _obtain_topology(master_address, cluster_def): 168 """Obtains TPU fabric topology.""" 169 try: 170 logging.info('Initializing TPU system (master: %s) to fetch topology ' 171 'for model parallelism. This might take a while.', 172 master_address) 173 with ops.Graph().as_default(): 174 session_config = get_session_config_with_timeout( 175 _INITIAL_TPU_SYSTEM_TIMEOUT_IN_MS, cluster_def) 176 with session_lib.Session( 177 master_address, config=session_config) as sess: 178 topology = sess.run(tpu.initialize_system()) 179 return topology 180 except errors.DeadlineExceededError: 181 raise ValueError( 182 'Fail to initialize TPU system with master (%s). ' 183 'Please double check the TPU system is functional.' % ( 184 master_address)) 185 186 187def get_session_config_with_timeout(timeout_in_secs, cluster_def): 188 """Returns a session given a timeout and a cluster configuration.""" 189 config_proto = config_pb2.ConfigProto( 190 operation_timeout_in_ms=timeout_in_secs, cluster_def=cluster_def) 191 return config_proto 192 193 194def master_job(master, cluster_def): 195 """Returns the canonical job name to use to place TPU computations on. 196 197 Args: 198 master: A `string` representing the TensorFlow master to use. 199 cluster_def: A ClusterDef object describing the TPU cluster. 200 201 Returns: 202 A string containing the job name, or None if no job should be specified. 203 204 Raises: 205 ValueError: If the user needs to specify a tpu_job_name, because we are 206 unable to infer the job name automatically, or if the user-specified job 207 names are inappropriate. 208 """ 209 # If the user specifies the tpu_job_name, use that. 210 211 if master in _LOCAL_MASTERS: 212 return None 213 214 if (not cluster_def or not cluster_def.job): 215 return _DEFAULT_JOB_NAME 216 job_names = set(job.name for job in cluster_def.job) 217 if _DEFAULT_JOB_NAME in job_names: 218 # b/37868888 tracks allowing ClusterSpec propagation to reuse job names. 219 raise ValueError('Currently, tpu_worker is not an allowed job name.') 220 if len(job_names) == 1: 221 return cluster_def.job[0].name 222 if len(job_names) == 2: 223 if _DEFAULT_COORDINATOR_JOB_NAME in job_names: 224 job_names.remove(_DEFAULT_COORDINATOR_JOB_NAME) 225 return job_names.pop() 226 # TODO(b/67716447): Include more sophisticated heuristics. 227 raise ValueError('Could not infer TPU job name.') 228