05 Dec 2023 - pinaki
In this blog, I will explain how to use docker in a computer that already has docker installed. This should be fairly quick read for people with less time. For inquisitive minds, I will also explain how to setup Docker in a computer for others to use. Finally, I shall explain about a variation of Docker popular in HPC, known as Singularity. This tutorial is focused towards users of docker and not developers.
Think of Docker as a better version of your Anaconda or Miniconda. It wraps all your libraries (python, C++, etc.) into a Docker Image (this is usually a folder or one file). This docker image is equivalent to your conda environment. You can create a seperate docker image for each of your projects. For AI/ML projects using PyTorch+cuda a docker image usually takes approx. 13GB disk space. You can create a docker image in your laptop (or some desktop used for debugging). Then transfer this docker image to any GPU server (HPC) and it will run without any issues.
The benefit of using docker is, previously, we created a miniconda environment in our local desktop. Then, to run our code for 2/3 days in a GPU server, we had to re-install the packages (especially some packages were build using cuda-11.7 but your server had cuda-11.3, etc.)
Remember, I am no expert on docker and it always best to learn from the people who developed it. So, if you can, learn from the official docs here: docker get-started
I shall assume you already have docker installed in your debugging GPU machine and you are using this machine from your laptop using vscode. You do not need docker installed in your laptop, you only need it in your debugging GPU machine. If you do not have docker installed in your debugging GPU machine, go to https://docs.docker.com/engine/install/ubuntu/.
Step-1: Pull an pre-built docker image from hub.docker.com.
I shall pull a docker with PyTorch2.0 and cuda 11.7 that you can find in this url: https://hub.docker.com/r/pytorch/pytorch/tags. Go to your debugging GPU desktop, open a terminal and type:
$ docker pull pytorch/pytorch:2.0.1-cuda11.7-cudnn8-runtime
Step-2: Join your debugging GPU using vscode and install a bunch of vscode extensions: Dev Containers, Docker, Remote - SSH, Remote - SSH: Editing Configurations Files, and Remote Explorer.
Step-3: (1) Start your docker image, (2) Attach your codebase, (3) Attach your local GPUs. Open the terminal inside your vscode and type the following commands:
$ docker image ls
REPOSITORY TAG IMAGE ID CREATED SIZE
pytorch/pytorch 2.0.0-cuda11.7-cudnn8-runtime 372afdac2915 8 months ago 6.44GB
pytorch/pytorch 1.12.1-cuda11.3-cudnn8-runtime 385dc33ab519 16 months ago 5.93GB
pytorch/pytorch 1.9.0-cuda11.1-cudnn8-runtime 12da3f8ec2ec 2 years ago 7.92GB
$ git clone https://github.com/pytorch/examples # Download dummy code examples to check your installation
$ export IMAGE_ID=372afdac2915 # image ID of your docker image
$ export PATH_TO_CODEBASE=$PWD/examples # absolute path to your codebase
$ export CONTAINER_NAME=test_docker # when you start a docker image, a container is created, which you will run. Name that container.
$ docker run -it -v $PATH_TO_CODEBASE:/workspace --gpus all --name test_docker $IMAGE_ID
root@0524673faa63:/workspace# ls
CODEOWNERS LICENSE dcgan fast_neural_style gcn mnist mnist_rnn run_python_examples.sh super_resolution vision_transformer
CODE_OF_CONDUCT.md README.md distributed fx imagenet mnist_forward_forward regression runtime.txt time_sequence_prediction word_language_model
CONTRIBUTING.md cpp docs gat legacy mnist_hogwild reinforcement_learning siamese_network vae
root@0524673faa63:/workspace# cd mnist
root@0524673faa63:/workspace/mnist# ls
README.md main.py requirements.txt
root@0524673faa63:/workspace/mnist# python main.py
Downloading http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz to ../data/MNIST/raw/train-images-idx3-ubyte.gz
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 9912422/9912422 [00:00<00:00, 20142812.27it/s]
Extracting ../data/MNIST/raw/train-images-idx3-ubyte.gz to ../data/MNIST/raw
Downloading http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz to ../data/MNIST/raw/train-labels-idx1-ubyte.gz
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 28881/28881 [00:00<00:00, 93181302.94it/s]
Extracting ../data/MNIST/raw/train-labels-idx1-ubyte.gz to ../data/MNIST/raw
...
...
Step-4: In step-3, we executed mnist example using terminal. Now, we do the same using vscode as shown in the following video:
Follow the official tutorial here: https://docs.docker.com/engine/install/ubuntu/
The steps in the official tutorial is pretty straight-forward. In case you are stuck, send me an email and I will populate the steps.
Comment
byu/elbigdoggogoog from discussion
indocker
So let’s get Singularity. To install you can follow the official docs here at: https://docs.sylabs.io/guides/latest/user-guide/quick_start.html
I will write the steps for Ubuntu. Run the following commands in terminal:
# Ensure repositories are up-to-date
$ sudo apt-get update
# Install debian packages for dependencies
$ sudo apt-get install -y \
autoconf \
automake \
cryptsetup \
git \
libfuse-dev \
libglib2.0-dev \
libseccomp-dev \
libtool \
pkg-config \
runc \
squashfs-tools \
squashfs-tools-ng \
uidmap \
wget \
zlib1g-dev
# Install Go
$ wget https://go.dev/dl/go1.21.5.linux-amd64.tar.gz
$ rm -rf /usr/local/go && tar -C /usr/local -xzf go1.21.5.linux-amd64.tar.gz
$ sudo echo "export PATH=$PATH:/usr/local/go/bin" >> $HOME/.profile
# Check Go version
$ go version
# Install Singularity
$ export VERSION=4.0.0 && \
wget https://github.com/sylabs/singularity/releases/download/v${VERSION}/singularity-ce-${VERSION}.tar.gz && \
tar -xzf singularity-ce-${VERSION}.tar.gz && \
cd singularity-ce-${VERSION}
$ ./mconfig && \
make -C builddir && \
sudo make -C builddir install
# Check if Singularity is installed
$ singularity help
Linux container platform optimized for High Performance Computing (HPC) and
Enterprise Performance Computing (EPC)
Usage:
singularity [global options...]
Now, we shall create a singularity container:
# Build Singularity Container
$ singularity build --sandbox test_singularity docker://pytorch/pytorch:2.0.0-cuda11.7-cudnn8-runtime
# Run the container
$ singularity shell --nv --writable test_singularity