Skip to content

Latest commit

 

History

History
58 lines (45 loc) · 2.51 KB

RESOURCE_GUIDE.md

File metadata and controls

58 lines (45 loc) · 2.51 KB
Raw

Resource Guide

This guide is meant to provide some resource, time, and money tradeoffs such that you can choose the best configuration for your data science experience!

STOPPING VM INSTANCES

Make sure to stop your VM instance when you are not using it!

When you start a VM, GCP will charge you an hourly rate (see below). If you are not actively running tasks on the VM or using the interface, you should stop the instance to save money. To do so, you can either do it through the web ui by clicking on the vertical dots at the right of your instance and stopping it, or use the command line from cloud shell:

gcloud compute instances stop [vm-instance-name]

GPU or No GPU

The NVIDIA Tesla K80 is significantly cheaper than the Tesla P100, but not nearly as powerful. The P100 can speed up deep network training by 10-20x compared to no GPU. If you plan to spend a lot of time visualizing and debugging data and less time running massive training jobs, it may be a good idea to start without a GPU. When you reach a step in the tutorial where you are training and optimizing deep neural nets is required, a GPU can saves you a lot of time, and potentially money. To compare costs and performance, consider the training task used in the cats project:

  • No GPU: A 2 CPU VM costs around $0.10 per hour when active. Training on the full dataset in a notebook (which only runs on 1 CPU) takes about 8 hours.
  • K80: around $0.55 per hour when active. Training takes about 31 minutes.
  • P100: around $1.50 per hour when active. Training takes about 25 minutes.

(It is possible that the default configuration does not optimize for P100, and further savings can be achieved.)

Dataflow CPUs

By default, our project is tuned to run tasks that will not incur more than a few dollars to your account, while ensuring that jobs complete quickly. If you would like to save money at the cost of extra time, dataflow jobs can be tuned using flags such as --num_workers and --worker_machine_type. The: example below dispatches a job on 20 machines with a worker type that contains 4 cpus per machine:

--num_workers 20 \
--worker_machine_type n1-standard-4 \

If you are running a lightweight task, such as a small sample of cat images, consider using fewer machines and/or changing the worker machine type to a simpler worker. This can reduce resources required to boot up and setup machines, and reduce communication overhead between machines.