CloudML is a managed version of Tensorflow run by Google. Instead of running TF by yourself you can simply use CloudML and forget all the issues related to infrastructure and scalability.
Running Distributed TensorFlow on Google CloudML
Getting ready
Here we assume that you have already created a Cloud Platform Project, enabling billing for your project, and enable the Google Compute Engine and the Cloud Machine Learning APIs. These steps are similar to the ones described in the previous recipes. This recipe is inspired by the MNIST training code available in https://cloud.google.com/ml-engine/docs/distributed-tensorflow-mnist-cloud-datalab.