An Example Using Preemptible VM for TensorFlow

This is an example to use preemptible VM for machine learning.

Requirements

• Google Cloud SDK
• Docker

Architecture

Key Points

• Use managed instance group with size one
  • Managed instance group can automatically restart preemptible VM
  • See the official document for details
• Make your instance to run Docker container when the instance starts
  • Container-Optimized OS is used in this example
  • Another choice is using startup script if you don't need Docker container
• Make instance group size zero after training finished

Run on Cloud

Configure Google Cloud SDK

gcloud config set account <your.google.account@gmail.com>
gcloud config set project <Your GCP Project>
gcloud auth configure-docker

Build Docker Image

./bin/docker_build.sh

Push Docker Image to Google Container Registry

./bin/docker_push.sh

Create Instance Template and Managed Instance Group

./bin/create_instance_group.sh

Run on Local

Build Docker Image

./bin/docker_build.sh

Run Docker Image

./bin/docker_run.sh