I start and terminate instances often for software development. This is a tool which makes it easy and quick to launch an instance configured and provisioned so you can start using it right away without manual configuration.

A nice feature is that it will use the ephemeral drives avaiable on the instance and create a raid volume /dev/md0 which will be mounted in /home. Many instance types have very fast local storage such as SSD disks and NVMe drives. Having the home folder in a raid0 of ephemeral drives makes working with the storage much faster than the default EBS root volume. This is done by userdata.py.

You would need Python3, boto3 and ansible installed. You can install them with the following commands:

pip3 install -r requirements.txt

You can copy the files that you want into homedir/ which will populate your home in the instance.

You should configure your AWS credentials and have EC2 permissions:

$ ./launch.py
instance_name [worker-piotr]:
instance_type [m1.xlarge]:
(public) ssh_key_file [/home/piotr/.ssh/id_rsa.pub]:
ami [ami-58d7e821]:
user name [piotr]:
[...]

After a few minutes the instance is created and provisioned and the hostname is printed on the output so you can ssh to it.

You can choose your ami depending on your region. For example using the ubuntu ami locator.

For gpu instances you can run nvidia_setup.sh once. Then you can check that the GPU is accessible from docker running:

nvidia-docker run -ti nvidia/cuda:9.1-cudnn7-devel nvidia-smi

Now you are ready to install your deep learning framework of choice.

The launch.py script is just a bunch of boto3 calls which configure the security groups, ssh key and instance. The initialization is done with cloud init using the cloud-init file and userdata.py which configures the raid of ephemerals.

After this is done, ansible is used to provision the instance by installing software such as docker etc.

Feel free to improve and modify the scripts and send any changes as a PR.