TonY

TonY is a framework to natively run deep learning jobs on Apache Hadoop. It currently supports TensorFlow, PyTorch, MXNet and Horovod. TonY enables running either single node or distributed training as a Hadoop application. This native connector, together with other TonY features, aims to run machine learning jobs reliably and flexibly. For a quick overview of TonY and comparisons to other frameworks, please see this presentation.

Compatibility Notes

TonY itself is compatible with Hadoop 2.6.0 (CDH5.11.0) and above. If you need GPU isolation from TonY, you need Hadoop 2.10 or higher for Hadoop 2, or Hadoop 3.1.0 or higher for Hadoop 3.

Build

TonY is built using Gradle. To build TonY, run:

./gradlew build

This will automatically run tests, if want to build without running tests, run:

./gradlew build -x test

The jar required to run TonY will be located in ./tony-cli/build/libs/.

Usage

There are two ways to launch your deep learning jobs with TonY:

• Use a zipped Python virtual environment.
• Use Docker container.

Use a zipped Python virtual environment

The difference between this approach and the one with Docker is

• You don't need to set up your Hadoop cluster with Docker support.
• There is no requirement on a Docker image registry.

As you know, nothing comes for free. If you don't want to bother setting your cluster with Docker support, you'd need to prepare a zipped virtual environment for your job and your cluster should have the same OS version as the computer which builds the Python virtual environment.

Python virtual environment in a zip

$ unzip -Z1 my-venv.zip | head -n 10
  Python/
  Python/bin/
  Python/bin/rst2xml.py
  Python/bin/wheel
  Python/bin/rst2html5.py
  Python/bin/rst2odt.py
  Python/bin/rst2s5.py
  Python/bin/pip2.7
  Python/bin/saved_model_cli
  Python/bin/rst2pseudoxml.pyc

TonY jar and tony.xml

MyJob/
  > src/
    > models/
      mnist_distributed.py
  tony.xml
  tony-cli-0.4.7-all.jar
  my-venv.zip # The additional file you need.

A similar tony.xml but without Docker related configurations:

$ cat tony/tony.xml
<configuration>
  <property>
    <name>tony.worker.instances</name>
    <value>4</value>
  </property>
  <property>
    <name>tony.worker.memory</name>
    <value>4g</value>
  </property>
  <property>
    <name>tony.worker.gpus</name>
    <value>1</value>
  </property>
  <property>
    <name>tony.ps.memory</name>
    <value>3g</value>
  </property>
</configuration>

Then you can launch your job:

$ java -cp "`hadoop classpath --glob`:MyJob/*:MyJob" \
            com.linkedin.tony.cli.ClusterSubmitter \
            -executes models/mnist_distributed.py \ # relative path to model program inside the src_dir
            -task_params '--input_dir /path/to/hdfs/input --output_dir /path/to/hdfs/output \
            -python_venv my-venv.zip \
            -python_binary_path Python/bin/python \  # relative path to the Python binary inside the my-venv.zip
            -src_dir src

Use a Docker container

Note that this requires you have a properly configured Hadoop cluster with Docker support. Check this documentation if you are unsure how to set it up. Assuming you have properly set up your Hadoop cluster with Docker container runtime, you should have already built a proper Docker image with required Hadoop configurations. The next thing you need is to install your Python dependencies inside your Docker image - TensorFlow or PyTorch.

Below is a folder structure of what you need to launch the job:

MyJob/
  > src/
    > models/
      mnist_distributed.py
  tony.xml
  tony-cli-0.4.7-all.jar

The src/ folder would contain all your training script. The tony.xml is used to config your training job. Specifically for using Docker as the container runtime, your configuration should be similar to something below:

$ cat MyJob/tony.xml
<configuration>
  <property>
    <name>tony.worker.instances</name>
    <value>4</value>
  </property>
  <property>
    <name>tony.worker.memory</name>
    <value>4g</value>
  </property>
  <property>
    <name>tony.worker.gpus</name>
    <value>1</value>
  </property>
  <property>
    <name>tony.ps.memory</name>
    <value>3g</value>
  </property>
  <property>
    <name>tony.docker.enabled</name>
    <value>true</value>
  </property>
  <property>
    <name>tony.docker.containers.image</name>
    <value>YOUR_DOCKER_IMAGE_NAME</value>
  </property>
</configuration>

For a full list of configurations, please see the wiki.

Now you're ready to launch your job:

$ java -cp "`hadoop classpath --glob`:MyJob/*:MyJob/" \
        com.linkedin.tony.cli.ClusterSubmitter \
        -executes models/mnist_distributed.py \
        -task_params '--input_dir /path/to/hdfs/input --output_dir /path/to/hdfs/output' \
        -src_dir src \
        -python_binary_path /home/user_name/python_virtual_env/bin/python

TonY arguments

The command line arguments are as follows:

TonY configurations

There are multiple ways to specify configurations for your TonY job. As above, you can create an XML file called tony.xml and add its parent directory to your java classpath.

Alternatively, you can pass -conf_file <name_of_conf_file> to the java command line if you have a file not named tony.xml containing your configurations. (As before, the parent directory of this file must be added to the java classpath.)

If you wish to override configurations from your configuration file via command line, you can do so by passing -conf <tony.conf.key>=<tony.conf.value> argument pairs on the command line.

Please check our wiki for all TonY configurations and their default values.

TonY Examples

Below are examples to run distributed deep learning jobs with TonY:

• Distributed MNIST with TensorFlow
• Distributed MNIST with PyTorch
• Distributed MNIST with Horovod
• Linear regression with MXNet
• TonY in Google Cloud Platform
• TonY in Azkaban video

More information

For more information about TonY, check out the following:

• TonY presentation at DataWorks Summit '19 in Washington, D.C.
• TonY OpML '19 paper
• TonY LinkedIn Engineering blog post

FAQ

1. My tensorflow process hangs with

   2018-09-13 03:02:31.538790: E tensorflow/core/distributed_runtime/master.cc:272] CreateSession failed because worker /job:worker/replica:0/task:0 returned error: Unavailable: OS Error
   INFO:tensorflow:An error was raised while a session was being created. This may be due to a preemption of a connected worker or parameter server. A new session will be created. Error: OS Error
   INFO:tensorflow:Graph was finalized.
   2018-09-13 03:03:33.792490: I tensorflow/core/distributed_runtime/master_session.cc:1150] Start master session ea811198d338cc1d with config:
   INFO:tensorflow:Waiting for model to be ready.  Ready_for_local_init_op:  Variables not initialized: conv1/Variable, conv1/Variable_1, conv2/Variable, conv2/Variable_1, fc1/Variable, fc1/Variable_1, fc2/Variable, fc2/Variable_1, global_step, adam_optimizer/beta1_power, adam_optimizer/beta2_power, conv1/Variable/Adam, conv1/Variable/Adam_1, conv1/Variable_1/Adam, conv1/Variable_1/Adam_1, conv2/Variable/Adam, conv2/Variable/Adam_1, conv2/Variable_1/Adam, conv2/Variable_1/Adam_1, fc1/Variable/Adam, fc1/Variable/Adam_1, fc1/Variable_1/Adam, fc1/Variable_1/Adam_1, fc2/Variable/Adam, fc2/Variable/Adam_1, fc2/Variable_1/Adam, fc2/Variable_1/Adam_1, ready: None

   Why?

   Try adding the path to your libjvm.so shared library to your LD_LIBRARY_PATH environment variable for your workers. See above for an example.

2. How do I configure arbitrary TensorFlow job types?

   Please see the wiki on TensorFlow task configuration for details.

3. My tensorflow's partial workers hang when chief finished. Or evaluator hang when chief and workers finished.

   Please see the PR#521 and PR#641 on Tensorflow configuration to solve it.