TensorFlowOnYARN

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

Goals

• 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

Architecture

Figure1. TOY Architecture

Features

• 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

0. Prepare the build environment following the instructions from https://www.tensorflow.org/install/install_sources

1. Clone the TensorFlowOnYARN repository.

   git clone --recursive https://github.com/Intel-bigdata/TensorFlowOnYARN

2. 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.

3. 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/mnist_feed.py \
     --ps_hosts="ps0.hostname:ps0.port,ps1.hostname:ps1.port" \
     --worker_hosts="worker0.hostname:worker0.port,worker1.hostname:worker1.port" \
     --task_index=0
   
   python examples/between-graph/mnist_feed.py \
     --ps_hosts="ps0.hostname:ps0.port,ps1.hostname:ps1.port" \
     --worker_hosts="worker0.hostname:worker0.port,worker1.hostname:worker1.port" \
     --task_index=1

4. To get ClusterSpec of an existing TensorFlow cluster launched by a previous YARN application.

   bin/ydl-tf cluster --app_id <Application ID>

5. 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>