- Create your AWS account
If you haven't got an account with Amazon AWS yet, you'll need to create one. Go to http://aws.amazon.com/, click Sign Up and follow the instructions. You'll probably have to receive a phone call from them to confirm your identity.
Create a new EC2 key pair and place your downloaded .pem file in your ~/.ssh directory.
Go to Security Credentials, click Access Keys, and Create New Access Key. Copy the secret access key and put it in a temporary file somewhere. You can only see it when you create it, but you can create as many as you want.
Define the following environment variables with the information you obtained in steps 2 and 3 (put these lines in ~/.bash_profile on OSX or ~/.bashrc on Linux):
export AWS_KEYPAIR_NAME=... # The name of the key pair you created in step 2
export AWS_SSH_KEY_FILENAME=... # The path where you put your .pem file in step 2 (e.g. ~/.ssh/foo.pem)
export AWS_ACCESS_KEY_ID=... # Found under Security Credentials -> Access Keys
export AWS_SECRET_ACCESS_KEY=... # The secret key you copied in step 3
Open a new terminal window or reload .bash_profile/.bashrc
- Install software
Install pip (a Python package manager) if you don't have it installed already.
Then (on the command line), install headintheclouds:
pip install headintheclouds
Download this repository:
git clone https://github.com/andreasjansson/simple-aws-gpu-setup.git
Cd into the repo root:
cd simple-aws-gpu-setup
If everything is installed properly, and all your environment variables from step 4 are correct, you should be able to type fab pricing and get a table of EC2 spot instance and hourly costs.
- Create your cheap GPU spot instance
On your terminal, type
fab ensemble.up:gpu
You should get a prompt like this:
Calculating changes...
The following servers will be created:
gpu
Do you wish to continue? [Y/n]
Answer Y
Now a spot instance will be created, based on the configuration in gpu.yaml. This will take around five minutes. If for some reason the current spot price is above the bid in gpu.yaml, edit that file such that bid is higher than the current asking price (or wait for a few hours for the asking price to go down again).
The image defined in gpu.yaml (ami-1ca6f874) is based on a stock Ubuntu 14.04 image (ami-8827efe0) and is bootstrapped with Theano, NumPy, SciPy, Emacs, and some other goodies. The actual bootstrap script is included in the bootstrap directory.
Once that's done you should see your server running if you type
fab nodes
- Write some Theano code
Log in to the new server
fab ssh
On the remote machine, start an IPython notebook on port 80
sudo ipython notebook --ip=0.0.0.0 --port=80 --no-browser
You can access the notebook in a web browser using the public_ip address from the fab nodes command.
Since this notebook is world readable, anyone can access your machine as root. That's a terrible idea. So shut down the notebook by hitting ctrl-c.
On your local machine, figure out your public IP address (either just google "what is my ip" or curl ifconfig.me). Replace 80: "*" in the firewall config of gpu.yaml with 80: YOUR_PUBLIC_IP. For the firewall to be applied, once again run
fab ensemble.up:gpu
Now you can safely start the IPython notebook again.
I added a check_theano notebook to the image. Open the IPython notebook web interface in a browser again and click the check_theano notebook. Once inside it, click the first cell and hit Run (or ctrl-return). It should say "Used the gpu".
Don't forget to back up your work regularly to S3 or some other persistent storage, since spot instances can be terminated at any point in time
When you're done with your instance you can terminate it by typing
fab -R gpu terminate
The scripts and files I used to build the image are in the bootstrap directory. To build with headintheclouds:
fab ensemble.up:build
fab -R build bootstrap
Or you can just copy the files over from bootstrap/files and run the bootstrap/bootstrap.sh script yourself.