Distributed TensorFlow

This document shows how to create a cluster of TensorFlow servers, and how to distribute a computation graph across that cluster. We assume that you are familiar with the basic concepts of writing TensorFlow programs.

Hello distributed TensorFlow!

To see a simple TensorFlow cluster in action, execute the following:

# Start a TensorFlow server as a single-process "cluster".
$ python
>>> import tensorflow as tf
>>> c = tf.constant("Hello, distributed TensorFlow!")
>>> server = tf.train.Server.create_local_server()
>>> sess = tf.Session(server.target)  # Create a session on the server.
>>> sess.run(c)
'Hello, distributed TensorFlow!'

The tf.train.Server.create_local_server method creates a single-process cluster, with an in-process server.

Create a cluster

A TensorFlow "cluster" is a set of "tasks" that participate in the distributed execution of a TensorFlow graph. Each task is associated with a TensorFlow "server", which contains a "master" that can be used to create sessions, and a "worker" that executes operations in the graph. A cluster can also be divided into one or more "jobs", where each job contains one or more tasks.

To create a cluster, you start one TensorFlow server per task in the cluster. Each task typically runs on a different machine, but you can run multiple tasks on the same machine (e.g. to control different GPU devices). In each task, do the following:

  1. Create a tf.train.ClusterSpec that describes all of the tasks in the cluster. This should be the same for each task.

  2. Create a tf.train.Server, passing the tf.train.ClusterSpec to the constructor, and identifying the local task with a job name and task index.

Create a tf.train.ClusterSpec to describe the cluster

The cluster specification dictionary maps job names to lists of network addresses. Pass this dictionary to the tf.train.ClusterSpec constructor. For example:

tf.train.ClusterSpec constructionAvailable tasks
tf.train.ClusterSpec({"local": ["localhost:2222", "localhost:2223"]})
/job:local/task:0
/job:local/task:1
tf.train.ClusterSpec({
    "worker": [
        "worker0.example.com:2222",
        "worker1.example.com:2222",
        "worker2.example.com:2222"
    ],
    "ps": [
        "ps0.example.com:2222",
        "ps1.example.com:2222"
    ]})
/job:worker/task:0
/job:worker/task:1
/job:worker/task:2
/job:ps/task:0
/job:ps/task:1

Create a tf.train.Server instance in each task

A tf.train.Server object contains a set of local devices, a set of connections to other tasks in its tf.train.ClusterSpec, and a tf.Session that can use these to perform a distributed computation. Each server is a member of a specific named job and has a task index within that job. A server can communicate with any other server in the cluster.

For example, to launch a cluster with two servers running on localhost:2222 and localhost:2223, run the following snippets in two different processes on the local machine:

# In task 0:
cluster = tf.train.ClusterSpec({"local": ["localhost:2222", "localhost:2223"]})
server = tf.train.Server(cluster, job_name="local", task_index=0)
# In task 1:
cluster = tf.train.ClusterSpec({"local": ["localhost:2222", "localhost:2223"]})
server = tf.train.Server(cluster, job_name="local", task_index=1)

Note: Manually specifying these cluster specifications can be tedious, especially for large clusters. We are working on tools for launching tasks programmatically, e.g. using a cluster manager like Kubernetes. If there are particular cluster managers for which you'd like to see support, please raise a GitHub issue.

Specifying distributed devices in your model

To place operations on a particular process, you can use the same tf.device function that is used to specify whether ops run on the CPU or GPU. For example:

with tf.device("/job:ps/task:0"):
  weights_1 = tf.Variable(...)
  biases_1 = tf.Variable(...)

with tf.device("/job:ps/task:1"):
  weights_2 = tf.Variable(...)
  biases_2 = tf.Variable(...)

with tf.device("/job:worker/task:7"):
  input, labels = ...
  layer_1 = tf.nn.relu(tf.matmul(input, weights_1) + biases_1)
  logits = tf.nn.relu(tf.matmul(layer_1, weights_2) + biases_2)
  # ...
  train_op = ...

with tf.Session("grpc://worker7.example.com:2222") as sess:
  for _ in range(10000):
    sess.run(train_op)

In the above example, the variables are created on two tasks in the ps job, and the compute-intensive part of the model is created in the worker job. TensorFlow will insert the appropriate data transfers between the jobs (from ps to worker for the forward pass, and from worker to ps for applying gradients).

Replicated training

A common training configuration, called "data parallelism," involves multiple tasks in a worker job training the same model on different mini-batches of data, updating shared parameters hosted in one or more tasks in a ps job. All tasks typically run on different machines. There are many ways to specify this structure in TensorFlow, and we are building libraries that will simplify the work of specifying a replicated model. Possible approaches include:

  • In-graph replication. In this approach, the client builds a single tf.Graph that contains one set of parameters (in tf.Variable nodes pinned to /job:ps); and multiple copies of the compute-intensive part of the model, each pinned to a different task in /job:worker.

  • Between-graph replication. In this approach, there is a separate client for each /job:worker task, typically in the same process as the worker task. Each client builds a similar graph containing the parameters (pinned to /job:ps as before using tf.train.replica_device_setter to map them deterministically to the same tasks); and a single copy of the compute-intensive part of the model, pinned to the local task in /job:worker.

  • Asynchronous training. In this approach, each replica of the graph has an independent training loop that executes without coordination. It is compatible with both forms of replication above.

  • Synchronous training. In this approach, all of the replicas read the same values for the current parameters, compute gradients in parallel, and then apply them together. It is compatible with in-graph replication (e.g. using gradient averaging as in the CIFAR-10 multi-GPU trainer), and between-graph replication (e.g. using the tf.train.SyncReplicasOptimizer).

Putting it all together: example trainer program

The following code shows the skeleton of a distributed trainer program, implementing between-graph replication and asynchronous training. It includes the code for the parameter server and worker tasks.

import argparse
import sys

import tensorflow as tf

FLAGS = None

def main(_):
  ps_hosts = FLAGS.ps_hosts.split(",")
  worker_hosts = FLAGS.worker_hosts.split(",")

  # Create a cluster from the parameter server and worker hosts.
  cluster = tf.train.ClusterSpec({"ps": ps_hosts, "worker": worker_hosts})

  # Create and start a server for the local task.
  server = tf.train.Server(cluster,
                           job_name=FLAGS.job_name,
                           task_index=FLAGS.task_index)

  if FLAGS.job_name == "ps":
    server.join()
  elif FLAGS.job_name == "worker":

    # Assigns ops to the local worker by default.
    with tf.device(tf.train.replica_device_setter(
        worker_device="/job:worker/task:%d" % FLAGS.task_index,
        cluster=cluster)):

      # Build model...
      loss = ...
      global_step = tf.contrib.framework.get_or_create_global_step()

      train_op = tf.train.AdagradOptimizer(0.01).minimize(
          loss, global_step=global_step)

    # The StopAtStepHook handles stopping after running given steps.
    hooks=[tf.train.StopAtStepHook(last_step=1000000)]

    # The MonitoredTrainingSession takes care of session initialization,
    # restoring from a checkpoint, saving to a checkpoint, and closing when done
    # or an error occurs.
    with tf.train.MonitoredTrainingSession(master=server.target,
                                           is_chief=(FLAGS.task_index == 0),
                                           checkpoint_dir="/tmp/train_logs",
                                           hooks=hooks) as mon_sess:
      while not mon_sess.should_stop():
        # Run a training step asynchronously.
        # See `tf.train.SyncReplicasOptimizer` for additional details on how to
        # perform *synchronous* training.
        # mon_sess.run handles AbortedError in case of preempted PS.
        mon_sess.run(train_op)

if __name__ == "__main__":
  parser = argparse.ArgumentParser()
  parser.register("type", "bool", lambda v: v.lower() == "true")
  # Flags for defining the tf.train.ClusterSpec
  parser.add_argument(
      "--ps_hosts",
      type=str,
      default="",
      help="Comma-separated list of hostname:port pairs"
  )
  parser.add_argument(
      "--worker_hosts",
      type=str,
      default="",
      help="Comma-separated list of hostname:port pairs"
  )
  parser.add_argument(
      "--job_name",
      type=str,
      default="",
      help="One of 'ps', 'worker'"
  )
  # Flags for defining the tf.train.Server
  parser.add_argument(
      "--task_index",
      type=int,
      default=0,
      help="Index of task within the job"
  )
  FLAGS, unparsed = parser.parse_known_args()
  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)

To start the trainer with two parameter servers and two workers, use the following command line (assuming the script is called trainer.py):

# On ps0.example.com:
$ python trainer.py \
     --ps_hosts=ps0.example.com:2222,ps1.example.com:2222 \
     --worker_hosts=worker0.example.com:2222,worker1.example.com:2222 \
     --job_name=ps --task_index=0
# On ps1.example.com:
$ python trainer.py \
     --ps_hosts=ps0.example.com:2222,ps1.example.com:2222 \
     --worker_hosts=worker0.example.com:2222,worker1.example.com:2222 \
     --job_name=ps --task_index=1
# On worker0.example.com:
$ python trainer.py \
     --ps_hosts=ps0.example.com:2222,ps1.example.com:2222 \
     --worker_hosts=worker0.example.com:2222,worker1.example.com:2222 \
     --job_name=worker --task_index=0
# On worker1.example.com:
$ python trainer.py \
     --ps_hosts=ps0.example.com:2222,ps1.example.com:2222 \
     --worker_hosts=worker0.example.com:2222,worker1.example.com:2222 \
     --job_name=worker --task_index=1

Glossary

Client

A client is typically a program that builds a TensorFlow graph and constructs a tensorflow::Session to interact with a cluster. Clients are typically written in Python or C++. A single client process can directly interact with multiple TensorFlow servers (see "Replicated training" above), and a single server can serve multiple clients.

Cluster

A TensorFlow cluster comprises a one or more "jobs", each divided into lists of one or more "tasks". A cluster is typically dedicated to a particular high-level objective, such as training a neural network, using many machines in parallel. A cluster is defined by a tf.train.ClusterSpec object.

Job

A job comprises a list of "tasks", which typically serve a common purpose. For example, a job named ps (for "parameter server") typically hosts nodes that store and update variables; while a job named worker typically hosts stateless nodes that perform compute-intensive tasks. The tasks in a job typically run on different machines. The set of job roles is flexible: for example, a worker may maintain some state.

Master service

An RPC service that provides remote access to a set of distributed devices, and acts as a session target. The master service implements the tensorflow::Session interface, and is responsible for coordinating work across one or more "worker services". All TensorFlow servers implement the master service.

Task

A task corresponds to a specific TensorFlow server, and typically corresponds to a single process. A task belongs to a particular "job" and is identified by its index within that job's list of tasks.

TensorFlow server A process running a tf.train.Server instance, which is a member of a cluster, and exports a "master service" and "worker service".

Worker service

An RPC service that executes parts of a TensorFlow graph using its local devices. A worker service implements worker_service.proto. All TensorFlow servers implement the worker service.