TonY - TensorFlow on YARN

This initialization action installs the latest version of TonY on a master node within a Google Cloud Dataproc cluster.

Using this initialization action

⚠️ NOTICE: See best practices of using initialization actions in production.

You can use this initialization action to create a new Dataproc cluster with TonY installed:

1. Use the gcloud command to create a new cluster with this initialization action.

   REGION=<region>
   CLUSTER_NAME=<cluster_name>
   gcloud dataproc clusters create ${CLUSTER_NAME} \
       --region ${REGION} \
       --initialization-actions gs://goog-dataproc-initialization-actions-${REGION}/tony/tony.sh

   You can pass specific metadata:

   REGION=<region>
   CLUSTER_NAME=<cluster_name>
   gcloud dataproc clusters create ${CLUSTER_NAME} \
       --region ${REGION} \
       --initialization-actions gs://goog-dataproc-initialization-actions-${REGION}/tony/tony.sh \
       --metadata name1=value1,name2=value2... 

   You can also create a cluster with GPUs:

   REGION=<region>
   CLUSTER_NAME=<cluster_name>
   gcloud dataproc clusters create ${CLUSTER_NAME} \
       --region ${REGION} \
       --master-aceelerator=nvidia-tesla-v100 \
       --worker-aceelerator=nvidia-tesla-v100 \
       --initialization-actions gs://goog-dataproc-initialization-actions-${REGION}/gpu/install_gpu_driver.sh,gs://goog-dataproc-initialization-actions-${REGION}/tony/tony.sh \
       --metadata gpu-driver-provider=NVIDIA \
       --metadata name1=value1,name2=value2... 

   Supported metadata parameters:

   • worker_instances
   • worker_memory
   • ps_instances
   • ps_memory
   • tensorflow version
   • tf_gpu

   These parameters are defined here: TonY configurations

   Example:

   REGION=<region>
   CLUSTER_NAME=<cluster_name>
   gcloud dataproc clusters create ${CLUSTER_NAME} \
       --region ${REGION} \
       --initialization-actions gs://goog-dataproc-initialization-actions-${REGION}/tony/tony.sh \
       --metadata worker_instances=4,worker_memory=4g,ps_instances=1,ps_memory=2g

   Note: For settings not defined in this configuration, you can pass a separate configuration when launching tasks or contribute to this repository and add support for more TonY configurations.

2. Once the cluster has been created, you can access the Hadoop web interface on the master node in a Dataproc cluster. To connect to the Hadoop web interface, you will need to create an SSH tunnel and use a SOCKS 5 Proxy with your web browser as described in the dataproc web interfaces documentation. In the opened web browser go to 'localhost:8088' and you should see the TonY UI.

3. TonY installation is located by default in the following folder:

   /opt/tony/TonY

4. TonY examples is located in the following folder:

   /opt/tony/TonY-samples

   By default we install examples for Tensorflow.

   • TensorFlow

For more information and to run some TonY examples, take a look at TonY examples A working example can be found on validate.sh

Testing

You can easily test TonY is running by using validate.sh ${cluster_prefix} ${bucket_name} script.

• cluster_prefix argument sets a prefix in created cluster name
• bucket_name argument sets a place where init action will be placed

This script is used for testing TonY using 1.5 Dataproc images with standard configurations. After clusters are created, script submits Hadoop jobs on them.

Important notes

• This script will install TonY in the master node only.
• Virtual environments are installed for only for TensorFlow examples.
• TonY is supported with STANDARD configuration (1 master/2+ workers) in Dataproc 1.5+.
• TonY supports GPU using Hadoop 3.1 (YARN-6223). To access this, you must use Dataproc 2.0+.