AWS EC2 ML instance setup

Steps for setting up an AWS EC2 instance for Docker+R+Keras

Based loosely on: https://github.com/NVIDIA/nvidia-docker (docker setup)

Start by launching a p2.xlarge instance with Ubuntu 16.04. Then follow the steps in each section below (via ssh into the instance, using the key pair you selected at launch).

Set up Docker

Original instructions: https://docs.docker.com/install/linux/docker-ce/ubuntu/

1. sudo apt-get update
2. sudo apt-get install apt-transport-https ca-certificates curl software-properties-common
3. curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo apt-key add -
4. sudo add-apt-repository "deb [arch=amd64] https://download.docker.com/linux/ubuntu $(lsb_release -cs) stable"
5. sudo apt-get update
6. sudo apt-get install -y docker-ce
7. sudo usermod -aG docker $USER
8. sudo reboot

(wait for reboot...usually takes 30-60 seconds)

Test: $ docker run hello-world

Install NVIDIA drivers

Original instructions: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/install-nvidia-driver.html#obtain-nvidia-driver-linux

1. sudo apt-get upgrade -y linux-aws
2. sudo reboot

(wait for reboot)

1. sudo apt-get install -y gcc make linux-headers-$(uname -r)
2. cd /tmp
3. curl -O http://us.download.nvidia.com/tesla/384.145/NVIDIA-Linux-x86_64-384.145.run
4. sudo /bin/sh ./NVIDIA-Linux-x86_64-384.145.run
5. rm NVIDIA-Linux-x86_64-384.145.run
6. sudo apt-get remove --purge -y gcc make linux-headers-$(uname -r)
7. sudo apt-get autoremove -y
8. sudo apt-get autoclean -y
9. sudo reboot

(wait for reboot)

Test driver install: $ nvidia-smi

Setup NVIDIA Docker runtime

Original instructions: https://github.com/NVIDIA/nvidia-docker

1. curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
2. distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
3. curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list
4. sudo apt-get update
5. sudo apt-get install -y nvidia-docker2
6. sudo pkill -SIGHUP dockerd

Test: sudo nvidia-container-cli --load-kmods info And: docker run --rm --runtime=nvidia -ti nvidia/cuda

Note that to run containers with the GPU available, you have to use the nvidia runtime