TensorFlow on YARN (TonY) is a framework to natively run TensorFlow on Apache Hadoop. TonY enables running either single node or distributed TensorFlow training as a Hadoop application. This native connector, together with other TonY features, aims to run TensorFlow jobs reliably and flexibly.

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

./gradlew build

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

TonY is a library, so it is as simple as running a java program. First, copy your artifacts to the machine with Hadoop installed, where you plan on running TonY from. This includes:

$ 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

$ ls tony/
  tony-cli-0.1.0-SNAPSHOT-all.jar  tony.xml

In the tony directory there’s also a tony.xml which contains all of your TonY job configurations. For example:

$ 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>

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

$ ls src/models/ | grep mnist_distributed
  mnist_distributed.py

Then you can launch your job:

$ java -cp "`hadoop classpath --glob`:tony/*:tony" \
            com.linkedin.tony.cli.ClusterSubmitter \
            -executes src/models/mnist_distributed.py \
            -task_params '--input_dir /path/to/hdfs/input --output_dir /path/to/hdfs/output --steps 2500 --batch_size 64' \
            -python_venv my-venv.zip \
            -python_binary_path Python/bin/python \
            -src_dir src \
            -shell_env LD_LIBRARY_PATH=/usr/java/latest/jre/lib/amd64/server

The command line arguments are as follows:

• executes describes the location to the entry point of your training code.
• task_params describe the command line arguments which will be passed to your entry point.
• python_venv describes the name of the zip locally which will invoke your python script.
• python_binary_path describes the relative path in your python virtual environment which contains the python binary, or an absolute path to use a python binary already installed on all worker nodes.
• src_dir specifies the name of the root directory locally which contains all of your python model source code. This directory will be copied to all worker nodes.
• shell_env specifies key-value pairs for environment variables which will be set in your python worker/ps processes.

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.

Finally, please check tony-default.xml or the wiki for default values of each TonY configuration.

Here is a full example of configuring your TonY application:

$ 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>
</configuration>

$ java -cp "`hadoop classpath --glob`:tony/*:tony" com.linkedin.tony.cli.ClusterSubmitter \
            -task_params '--data_dir hdfs://default/data/mnist --working_dir hdfs://default/mnist/working_dir --steps 2500 --batch_size 64' \
            -python_binary_path Python/bin/python \
            -python_venv my-venv.zip \
            -executes src/mnist_distributed.py \
            -shell_env LD_LIBRARY_PATH=/usr/java/latest/jre/lib/amd64/server \
            -conf tony.ps.instances=2 \
            -conf tony.worker.instances=2

CLI configurations have highest priority, so we will get 2 ps instances and 2 worker instances. Then the XML file takes next priority so each worker will get 4g memory and 1 GPU. Finally every other configuration will be default value, e.g. each ps will get 2g memory.

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.