# Copyright 2017 Yahoo Inc.
# Licensed under the terms of the Apache 2.0 license.
# Please see LICENSE file in the project root for terms.
"""This module provides low-level functions for managing the TensorFlowOnSpark cluster."""
from __future__ import absolute_import
from __future__ import division
from __future__ import nested_scopes
from __future__ import print_function
import json
import logging
import multiprocessing
import os
import pkg_resources
import platform
import socket
import subprocess
import sys
import uuid
import time
import traceback
from packaging import version
from threading import Thread
from . import TFManager
from . import TFNode
from . import gpu_info
from . import marker
from . import reservation
from . import util
logger = logging.getLogger(__name__)
TF_VERSION = pkg_resources.get_distribution('tensorflow').version
def _has_spark_resource_api():
"""Returns true if Spark 3+ resource API is available"""
import pyspark
return version.parse(pyspark.__version__).base_version >= version.parse("3.0.0").base_version
def _get_cluster_spec(sorted_cluster_info):
"""Given a list of node metadata sorted by executor_id, returns a tensorflow cluster_spec"""
cluster_spec = {}
last_executor_id = -1
for node in sorted_cluster_info:
if (node['executor_id'] == last_executor_id):
raise Exception("Duplicate worker/task in cluster_info")
last_executor_id = node['executor_id']
logger.info("node: {0}".format(node))
(njob, nhost, nport) = (node['job_name'], node['host'], node['port'])
hosts = [] if njob not in cluster_spec else cluster_spec[njob]
hosts.append("{0}:{1}".format(nhost, nport))
cluster_spec[njob] = hosts
return cluster_spec
class TFNodeContext:
"""Encapsulates unique metadata for a TensorFlowOnSpark node/executor and provides methods to interact with Spark and HDFS.
An instance of this object will be passed to the TensorFlow "main" function via the `ctx` argument.
To simply the end-user API, this class now mirrors the functions of the TFNode module.
Args:
:executor_id: integer identifier for this executor, per ``nodeRDD = sc.parallelize(range(num_executors), num_executors).``
:job_name: TensorFlow job name (e.g. 'ps' or 'worker') of this TF node, per cluster_spec.
:task_index: integer rank per job_name, e.g. "worker:0", "worker:1", "ps:0".
:cluster_spec: dictionary for constructing a tf.train.ClusterSpec.
:defaultFS: string representation of default FileSystem, e.g. ``file://`` or ``hdfs://<namenode>:8020/``.
:working_dir: the current working directory for local filesystems, or YARN containers.
:mgr: TFManager instance for this Python worker.
"""
def __init__(self, executor_id=0, job_name='', task_index=0, cluster_spec={}, defaultFS='file://', working_dir='.', mgr=None):
self.worker_num = executor_id # for backwards-compatibility
self.executor_id = executor_id
self.job_name = job_name
self.task_index = task_index
self.cluster_spec = cluster_spec
self.num_workers = sum([len(v) for k, v in cluster_spec.items() if k == 'master' or k == 'chief' or k == 'worker'])
self.defaultFS = defaultFS
self.working_dir = working_dir
self.mgr = mgr
def absolute_path(self, path):
"""Convenience function to access ``TFNode.hdfs_path`` directly from this object instance."""
return TFNode.hdfs_path(self, path)
def start_cluster_server(self, num_gpus=1, rdma=False):
"""Convenience function to access ``TFNode.start_cluster_server`` directly from this object instance."""
return TFNode.start_cluster_server(self, num_gpus, rdma)
def export_saved_model(self, sess, export_dir, tag_set, signatures):
"""Convenience function to access ``TFNode.export_saved_model`` directly from this object instance."""
TFNode.export_saved_model(sess, export_dir, tag_set, signatures)
def get_data_feed(self, train_mode=True, qname_in='input', qname_out='output', input_mapping=None):
"""Convenience function to access ``TFNode.DataFeed`` directly from this object instance."""
return TFNode.DataFeed(self.mgr, train_mode, qname_in, qname_out, input_mapping)
class TFSparkNode(object):
"""Low-level functions used by the high-level TFCluster APIs to manage cluster state.
**This class is not intended for end-users (see TFNode for end-user APIs)**.
For cluster management, this wraps the per-node cluster logic as Spark RDD mapPartitions functions, where the RDD is expected to be
a "nodeRDD" of the form: ``nodeRDD = sc.parallelize(range(num_executors), num_executors)``.
For data feeding, this wraps the feeding logic as Spark RDD mapPartitions functions on a standard "dataRDD".
This also manages a reference to the TFManager "singleton" per executor. Since Spark can spawn more than one python-worker
per executor, this will reconnect to the "singleton" instance as needed.
"""
mgr = None #: TFManager instance
cluster_id = None #: Unique ID for a given TensorFlowOnSpark cluster, used for invalidating state for new clusters.
def _get_manager(cluster_info, host, executor_id):
"""Returns this executor's "singleton" instance of the multiprocessing.Manager, reconnecting per python-worker if needed.
Args:
:cluster_info: cluster node reservations
:host: host IP address
:executor_id: unique id per executor (created during initial call to run())
Returns:
TFManager instance for this executor/python-worker
"""
for node in cluster_info:
if node['host'] == host and node['executor_id'] == executor_id:
addr = node['addr']
authkey = node['authkey']
TFSparkNode.mgr = TFManager.connect(addr, authkey)
break
if TFSparkNode.mgr is None:
msg = "No TFManager found on this node, please ensure that:\n" + \
"1. Spark num_executors matches TensorFlow cluster_size\n" + \
"2. Spark tasks per executor is 1\n" + \
"3. Spark dynamic allocation is disabled\n" + \
"4. There are no other root-cause exceptions on other nodes\n"
raise Exception(msg)
logger.info("Connected to TFSparkNode.mgr on {0}, executor={1}, state={2}".format(host, executor_id, str(TFSparkNode.mgr.get('state'))))
return TFSparkNode.mgr
def run(fn, tf_args, cluster_meta, tensorboard, log_dir, queues, background):
"""Wraps the user-provided TensorFlow main function in a Spark mapPartitions function.
Args:
:fn: TensorFlow "main" function provided by the user.
:tf_args: ``argparse`` args, or command line ``ARGV``. These will be passed to the ``fn``.
:cluster_meta: dictionary of cluster metadata (e.g. cluster_id, reservation.Server address, etc).
:tensorboard: boolean indicating if the chief worker should spawn a Tensorboard server.
:log_dir: directory to save tensorboard event logs. If None, defaults to a fixed path on local filesystem.
:queues: *INTERNAL_USE*
:background: boolean indicating if the TensorFlow "main" function should be run in a background process.
Returns:
A nodeRDD.mapPartitions() function.
"""
def _mapfn(iter):
# Note: consuming the input iterator helps Pyspark re-use this worker,
for i in iter:
executor_id = i
def _get_gpus(cluster_spec=None):
gpus = []
is_k8s = 'SPARK_EXECUTOR_POD_IP' in os.environ
# handle explicitly configured tf_args.num_gpus
if 'num_gpus' in tf_args:
requested_gpus = tf_args.num_gpus
user_requested = True
else:
requested_gpus = 0
user_requested = False
# first, try Spark 3 resources API, returning all visible GPUs
# note: num_gpus arg is only used (if supplied) to limit/truncate visible devices
if _has_spark_resource_api():
from pyspark import TaskContext
context = TaskContext()
resources = context.resources()
if resources and 'gpu' in resources:
# get all GPUs assigned by resource manager
gpus = context.resources()['gpu'].addresses
logger.info("Spark gpu resources: {}".format(gpus))
if user_requested:
if requested_gpus < len(gpus):
# override/truncate list, if explicitly configured
logger.warn("Requested {} GPU(s), but {} available".format(requested_gpus, len(gpus)))
gpus = gpus[:requested_gpus]
else:
# implicitly requested by Spark 3
requested_gpus = len(gpus)
# if not in K8s pod and GPUs available, just use original allocation code (defaulting to 1 GPU if available)
# Note: for K8s, there is a bug with the Nvidia device_plugin which can show GPUs for non-GPU pods that are hosted on GPU nodes
if not is_k8s and gpu_info.is_gpu_available() and not gpus:
# default to one GPU if not specified explicitly
requested_gpus = max(1, requested_gpus) if not user_requested else requested_gpus
if requested_gpus > 0:
if cluster_spec:
# compute my index relative to other nodes on the same host (for GPU allocation)
my_addr = cluster_spec[job_name][task_index]
my_host = my_addr.split(':')[0]
flattened = [v for sublist in cluster_spec.values() for v in sublist]
local_peers = [p for p in flattened if p.startswith(my_host)]
my_index = local_peers.index(my_addr)
else:
my_index = 0
# try to allocate a GPU
gpus = gpu_info.get_gpus(requested_gpus, my_index, format=gpu_info.AS_LIST)
if user_requested and len(gpus) < requested_gpus:
raise Exception("Unable to allocate {} GPU(s) from available GPUs: {}".format(requested_gpus, gpus))
gpus_to_use = ','.join(gpus)
if gpus:
logger.info("Requested {} GPU(s), setting CUDA_VISIBLE_DEVICES={}".format(requested_gpus if user_requested else len(gpus), gpus_to_use))
os.environ['CUDA_VISIBLE_DEVICES'] = gpus_to_use
# try GPU allocation at executor startup so we can try to fail out if unsuccessful
_get_gpus()
# assign TF job/task based on provided cluster_spec template (or use default/null values)
job_name = 'default'
task_index = -1
cluster_id = cluster_meta['id']
cluster_template = cluster_meta['cluster_template']
for jobtype in cluster_template:
nodes = cluster_template[jobtype]
if executor_id in nodes:
job_name = jobtype
task_index = nodes.index(executor_id)
break
# get unique key (hostname, executor_id) for this executor
host = util.get_ip_address()
util.write_executor_id(executor_id)
port = 0
# check for existing TFManagers
if TFSparkNode.mgr is not None and str(TFSparkNode.mgr.get('state')) != "'stopped'":
if TFSparkNode.cluster_id == cluster_id:
# raise an exception to force Spark to retry this "reservation" task on another executor
raise Exception("TFManager already started on {0}, executor={1}, state={2}".format(host, executor_id, str(TFSparkNode.mgr.get("state"))))
else:
# old state, just continue with creating new manager
logger.warn("Ignoring old TFManager with cluster_id {0}, requested cluster_id {1}".format(TFSparkNode.cluster_id, cluster_id))
# start a TFManager and get a free port
# use a random uuid as the authkey
authkey = uuid.uuid4().bytes
addr = None
if job_name in ('ps', 'evaluator'):
# PS nodes must be remotely accessible in order to shutdown from Spark driver.
TFSparkNode.mgr = TFManager.start(authkey, ['control', 'error'], 'remote')
addr = (host, TFSparkNode.mgr.address[1])
else:
# worker nodes only need to be locally accessible within the executor for data feeding
TFSparkNode.mgr = TFManager.start(authkey, queues)
addr = TFSparkNode.mgr.address
# initialize mgr state
TFSparkNode.mgr.set('state', 'running')
TFSparkNode.cluster_id = cluster_id
# expand Hadoop classpath wildcards for JNI (Spark 2.x)
if 'HADOOP_PREFIX' in os.environ:
classpath = os.environ['CLASSPATH']
hadoop_path = os.path.join(os.environ['HADOOP_PREFIX'], 'bin', 'hadoop')
hadoop_classpath = subprocess.check_output([hadoop_path, 'classpath', '--glob']).decode()
logger.debug("CLASSPATH: {0}".format(hadoop_classpath))
os.environ['CLASSPATH'] = classpath + os.pathsep + hadoop_classpath
# start TensorBoard if requested, on 'worker:0' if available (for backwards-compatibility), otherwise on 'chief:0' or 'master:0'
job_names = sorted([k for k in cluster_template.keys() if k in ['chief', 'master', 'worker']])
tb_job_name = 'worker' if 'worker' in job_names else job_names[0]
tb_pid = 0
tb_port = 0
if tensorboard and job_name == tb_job_name and task_index == 0:
if 'TENSORBOARD_PORT' in os.environ:
# use port defined in env var
tb_port = int(os.environ['TENSORBOARD_PORT'])
else:
# otherwise, find a free port
tb_sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
tb_sock.bind(('', 0))
tb_port = tb_sock.getsockname()[1]
tb_sock.close()
logdir = log_dir if log_dir else "tensorboard_%d" % executor_id
# search for tensorboard in python/bin, PATH, and PYTHONPATH
pypath = sys.executable
pydir = os.path.dirname(pypath)
sys_path = os.pathsep.join(sys.path)
search_path = os.pathsep.join([pydir, sys_path, os.environ['PATH'], os.environ['PYTHONPATH']])
tb_path = util.find_in_path(search_path, 'tensorboard') # executable in PATH
if not tb_path:
tb_path = util.find_in_path(search_path, 'tensorboard/main.py') # TF 1.3+
if not tb_path:
tb_path = util.find_in_path(search_path, 'tensorflow/tensorboard/__main__.py') # TF 1.2-
if not tb_path:
raise Exception("Unable to find 'tensorboard' in: {}".format(search_path))
# launch tensorboard
if version.parse(TF_VERSION) >= version.parse('2.0.0'):
tb_proc = subprocess.Popen([pypath, tb_path, "--reload_multifile=True", "--logdir=%s" % logdir, "--port=%d" % tb_port], env=os.environ)
else:
tb_proc = subprocess.Popen([pypath, tb_path, "--logdir=%s" % logdir, "--port=%d" % tb_port], env=os.environ)
tb_pid = tb_proc.pid
# check server to see if this task is being retried (i.e. already reserved)
client = reservation.Client(cluster_meta['server_addr'])
cluster_info = client.get_reservations()
tmp_sock = None
node_meta = None
for node in cluster_info:
(nhost, nexec) = (node['host'], node['executor_id'])
if nhost == host and nexec == executor_id:
node_meta = node
port = node['port']
# if not already done, register everything we need to set up the cluster
if node_meta is None:
if 'TENSORFLOW_PORT' in os.environ:
# use port defined in env var
port = int(os.environ['TENSORFLOW_PORT'])
else:
# otherwise, find a free port
tmp_sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
tmp_sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
tmp_sock.bind(('', port))
port = tmp_sock.getsockname()[1]
node_meta = {
'executor_id': executor_id,
'host': host,
'job_name': job_name,
'task_index': task_index,
'port': port,
'tb_pid': tb_pid,
'tb_port': tb_port,
'addr': addr,
'authkey': authkey
}
# register node metadata with server
logger.info("TFSparkNode.reserve: {0}".format(node_meta))
client.register(node_meta)
# wait for other nodes to finish reservations
cluster_info = client.await_reservations()
client.close()
# construct a TensorFlow clusterspec from cluster_info
sorted_cluster_info = sorted(cluster_info, key=lambda k: k['executor_id'])
cluster_spec = _get_cluster_spec(sorted_cluster_info)
# update TF_CONFIG if cluster spec has a 'master' node (i.e. tf.estimator)
if 'master' in cluster_spec or 'chief' in cluster_spec:
tf_config = json.dumps({
'cluster': cluster_spec,
'task': {'type': job_name, 'index': task_index},
'environment': 'cloud'
})
logger.info("export TF_CONFIG: {}".format(tf_config))
os.environ['TF_CONFIG'] = tf_config
# reserve GPU(s) again, just before launching TF process (in case situation has changed)
# and setup CUDA_VISIBLE_DEVICES accordingly
_get_gpus(cluster_spec=cluster_spec)
# create a context object to hold metadata for TF
ctx = TFNodeContext(executor_id, job_name, task_index, cluster_spec, cluster_meta['default_fs'], cluster_meta['working_dir'], TFSparkNode.mgr)
# release port reserved for TF as late as possible
if tmp_sock is not None:
tmp_sock.close()
# Background mode relies reuse of python worker in Spark.
if background:
# However, reuse of python worker can't work on Windows, we need to check if the current
# script runs on Windows or not.
if os.name == 'nt' or platform.system() == 'Windows':
raise Exception("Background mode is not supported on Windows.")
# Check if the config of reuse python worker is enabled on Spark.
if not os.environ.get("SPARK_REUSE_WORKER"):
raise Exception("Background mode relies reuse of python worker on Spark. This config 'spark.python.worker.reuse' is not enabled on Spark. Please enable it before using background.")
def wrapper_fn(args, context):
"""Wrapper function that sets the sys.argv of the executor."""
if isinstance(args, list):
sys.argv = args
fn(args, context)
def wrapper_fn_background(args, context):
"""Wrapper function that signals exceptions to foreground process."""
errq = TFSparkNode.mgr.get_queue('error')
try:
wrapper_fn(args, context)
except Exception:
errq.put(traceback.format_exc())
if job_name in ('ps', 'evaluator') or background:
# invoke the TensorFlow main function in a background thread
logger.info("Starting TensorFlow {0}:{1} as {2} on cluster node {3} on background process".format(
job_name, task_index, job_name, executor_id))
p = multiprocessing.Process(target=wrapper_fn_background, args=(tf_args, ctx))
if job_name in ('ps', 'evaluator'):
p.daemon = True
p.start()
# for ps and evaluator nodes, wait indefinitely in foreground thread for a "control" event (None == "stop")
if job_name in ('ps', 'evaluator'):
queue = TFSparkNode.mgr.get_queue('control')
equeue = TFSparkNode.mgr.get_queue('error')
done = False
while not done:
while (queue.empty() and equeue.empty()):
time.sleep(1)
if (not equeue.empty()):
e_str = equeue.get()
raise Exception("Exception in " + job_name + ":\n" + e_str)
msg = queue.get(block=True)
logger.info("Got msg: {0}".format(msg))
if msg is None:
logger.info("Terminating {}".format(job_name))
TFSparkNode.mgr.set('state', 'stopped')
done = True
queue.task_done()
else:
# otherwise, just run TF function in the main executor/worker thread
logger.info("Starting TensorFlow {0}:{1} on cluster node {2} on foreground thread".format(job_name, task_index, executor_id))
wrapper_fn(tf_args, ctx)
logger.info("Finished TensorFlow {0}:{1} on cluster node {2}".format(job_name, task_index, executor_id))
return _mapfn
def train(cluster_info, cluster_meta, feed_timeout=600, qname='input'):
"""Feeds Spark partitions into the shared multiprocessing.Queue.
Args:
:cluster_info: node reservation information for the cluster (e.g. host, executor_id, pid, ports, etc)
:cluster_meta: dictionary of cluster metadata (e.g. cluster_id, reservation.Server address, etc)
:feed_timeout: number of seconds after which data feeding times out (600 sec default)
:qname: *INTERNAL_USE*
Returns:
A dataRDD.mapPartitions() function
"""
def _train(iter):
# get shared queue, reconnecting if necessary
mgr = _get_manager(cluster_info, util.get_ip_address(), util.read_executor_id())
try:
queue = mgr.get_queue(qname)
equeue = mgr.get_queue('error')
except (AttributeError, KeyError):
msg = "Queue '{}' not found on this node, check for exceptions on other nodes.".format(qname)
raise Exception(msg)
state = str(mgr.get('state'))
logger.info("mgr.state={0}".format(state))
terminating = state == "'terminating'"
if terminating:
logger.info("mgr is terminating, skipping partition")
count = sum(1 for item in iter)
logger.info("Skipped {0} items from partition".format(count))
else:
logger.info("Feeding partition {0} into {1} queue {2}".format(iter, qname, queue))
count = 0
for item in iter:
count += 1
queue.put(item, block=True)
# wait for consumers to finish processing all items in queue before "finishing" this iterator
joinThr = Thread(target=queue.join)
joinThr.start()
timeout = feed_timeout
while (joinThr.isAlive()):
if (not equeue.empty()):
e_str = equeue.get()
raise Exception("Exception in worker:\n" + e_str)
time.sleep(1)
timeout -= 1
if timeout <= 0:
raise Exception("Timeout while feeding partition")
logger.info("Processed {0} items in partition".format(count))
# check if TF is terminating feed after this partition
if not terminating:
state = str(mgr.get('state'))
terminating = state == "'terminating'"
if terminating:
try:
logger.info("TFSparkNode: requesting stop")
client = reservation.Client(cluster_meta['server_addr'])
client.request_stop()
client.close()
except Exception as e:
# ignore any errors while requesting stop
logger.debug("Error while requesting stop: {0}".format(e))
return [terminating]
return _train
def inference(cluster_info, feed_timeout=600, qname='input'):
"""Feeds Spark partitions into the shared multiprocessing.Queue and returns inference results.
Args:
:cluster_info: node reservation information for the cluster (e.g. host, executor_id, pid, ports, etc)
:feed_timeout: number of seconds after which data feeding times out (600 sec default)
:qname: *INTERNAL_USE*
Returns:
A dataRDD.mapPartitions() function
"""
def _inference(iter):
# get shared queue, reconnecting if necessary
mgr = _get_manager(cluster_info, util.get_ip_address(), util.read_executor_id())
try:
queue_in = mgr.get_queue(qname)
equeue = mgr.get_queue('error')
except (AttributeError, KeyError):
msg = "Queue '{}' not found on this node, check for exceptions on other nodes.".format(qname)
raise Exception(msg)
logger.info("Feeding partition {0} into {1} queue {2}".format(iter, qname, queue_in))
count = 0
for item in iter:
count += 1
queue_in.put(item, block=True)
# signal "end of partition"
queue_in.put(marker.EndPartition())
# skip empty partitions
if count == 0:
return []
# wait for consumers to finish processing all items in queue before "finishing" this iterator
joinThr = Thread(target=queue_in.join)
joinThr.start()
timeout = feed_timeout
while (joinThr.isAlive()):
if (not equeue.empty()):
e_str = equeue.get()
raise Exception("Exception in worker:\n" + e_str)
time.sleep(1)
timeout -= 1
if timeout <= 0:
raise Exception("Timeout while feeding partition")
logger.info("Processed {0} items in partition".format(count))
# read result queue
results = []
queue_out = mgr.get_queue('output')
while count > 0:
result = queue_out.get(block=True)
results.append(result)
count -= 1
queue_out.task_done()
logger.info("Finished processing partition")
return results
return _inference
def shutdown(cluster_info, grace_secs=0, queues=['input']):
"""Stops all TensorFlow nodes by feeding ``None`` into the multiprocessing.Queues.
Args:
:cluster_info: node reservation information for the cluster (e.g. host, executor_id, pid, ports, etc).
:queues: *INTERNAL_USE*
Returns:
A nodeRDD.mapPartitions() function
"""
def _shutdown(iter):
host = util.get_ip_address()
executor_id = util.read_executor_id()
# reconnect to shared queue
mgr = _get_manager(cluster_info, host, executor_id)
# send SIGTERM to Tensorboard proc (if running)
for node in cluster_info:
if node['host'] == host and node['executor_id'] == executor_id:
tb_pid = node['tb_pid']
if tb_pid != 0:
logger.info("Stopping tensorboard (pid={0})".format(tb_pid))
subprocess.Popen(["kill", str(tb_pid)])
# terminate any listening queues
logger.info("Stopping all queues")
for q in queues:
if q != 'error':
try:
queue = mgr.get_queue(q)
logger.info("Feeding None into {0} queue".format(q))
queue.put(None, block=True)
except (AttributeError, KeyError):
msg = "Queue '{}' not found on this node, check for exceptions on other nodes.".format(q)
raise Exception(msg)
# wait for grace period (after terminating feed queues)
if grace_secs > 0:
logger.info("Waiting for {} second grace period".format(grace_secs))
time.sleep(grace_secs)
# then check for any late exceptions
equeue = mgr.get_queue('error')
if (not equeue.empty()):
# note: "peek" this queue, since otherwise Spark might retry this "failed" task, find no errors in queue, and finish the job with SUCCESS
e_str = equeue.get()
equeue.put(e_str)
raise Exception("Exception in worker:\n" + e_str)
logger.info("Setting mgr.state to 'stopped'")
mgr.set('state', 'stopped')
return [True]
return _shutdown
