Support TensorFlow on YARN
Switch branches/tags
Nothing to show
Clone or download
Latest commit f425999 Apr 18, 2017

TensorFlowOnYARN Build Status

TensorFlow on YARN (TOY) is a toolkit to enable Hadoop users an easy way to run TensorFlow applications in distributed pattern and accomplish tasks including model management and serving inference.

  • This project focuses on support of running Tensorflow on YARN, as part of Deep Learning on Hadoop (HDL) effort.
  • YARN-6043


  • Support all TensorFlow components on YARN, TensorFlow distributed cluster, TensorFlow serving, TensorBoard, etc.
  • Support multi-tenants with consideration of different types of users, such as devOp, data scientist and data engineer
  • Support running TensorFlow application in a short-time/long-running job manner of both between-graph mode and in-graph mode
  • Support model management to deploy and also support a service layer to handle upper layer's like Spark or web backend inference request easily
  • Minor or no changes required to run user’s existing TensorFlow application(can be written in all officially supported languages including Python, C++, Java and Go)

Note that current project is a prototype with limitation and is still under development


Figure1. TOY Architecture


  • Launch a TensorFlow cluster with specified number of worker and PS server
  • Replace python layer with java bridge layer to start server
  • Generate ClusterSpec dynamically
  • RPC support for client to get ClusterSpec from AM
  • Signal handling for graceful shutdown
  • Package TensorFlow runtime as a resource that can be distributed easily
  • Run in-graph TensorFlow application in client mode
  • TensorBoard support
  • Better handling of network port conflicts
  • Fault tolerance
  • Cluster mode based on Docker
  • Real-time logging support
  • Code refine and more tests

Quick Start

  1. Prepare the build environment following the instructions from

  2. Clone the TensorFlowOnYARN repository.

    git clone --recursive
  3. Build the assembly.

    cd TensorFlowOnYARN/tensorflow-parent
    mvn package -Pnative -Pdist

    tensorflow-yarn-${VERSION}.tar.gz and tensorflow-yarn-${VERSION}.zip are built out in the tensorflow-parent/tensorflow-yarn-dist/target directory. Distribute the assembly to the client node of a YARN cluster and extract.

  4. Run the between-graph mnist example.

    cd tensorflow-yarn-${VERSION}
    bin/ydl-tf launch --num_worker 2 --num_ps 2

    This will launch a YARN application, which creates a tf.train.Server instance for each task. A ClusterSpec is printed on the console such that you can submit the training script to. e.g.

    ClusterSpec: {"ps":["node1:22257","node2:22222"],"worker":["node3:22253","node2:22255"]}
    python examples/between-graph/ \
      --ps_hosts="ps0.hostname:ps0.port,ps1.hostname:ps1.port" \
      --worker_hosts="worker0.hostname:worker0.port,worker1.hostname:worker1.port" \
    python examples/between-graph/ \
      --ps_hosts="ps0.hostname:ps0.port,ps1.hostname:ps1.port" \
      --worker_hosts="worker0.hostname:worker0.port,worker1.hostname:worker1.port" \
  5. To get ClusterSpec of an existing TensorFlow cluster launched by a previous YARN application.

    bin/ydl-tf cluster --app_id <Application ID>
  6. You may also use YARN commands through ydl-tf.

    For example, to get running application list,

    bin/ydl-tf application --list

    or to kill an existing YARN application(TensorFlow cluster),

    bin/ydl-tf kill --application <Application ID>