Even for moderately complex projects, the size of the software dependency stack can be huge.
Take a simple pipeline to build a pdf report for an analysis scripted in R using
Rmarkdown, for example.
To make this reproducible, not only do (i) the respective R packages need to be installed and (ii) the R version needs to be the same, but also (iii) the versions of
LaTeX need to be the same as during runtime.
Instead of trying to resolve these dependencies via a package manager (such as conda) - which also depends on all required software being available in a single package manager - it might be easier to create a snapshot of the entire computing environment including all dependencies. These computing environments are then self-contained, hence the name ‘containers’.
This RedHat blog post provides an introduction to containers and container terminology.
What are Containers?#
Containers allow a researcher to package up a project with all of the parts it needs - such as libraries, dependencies, and system settings - and ship it all out as one package. Anyone can then open up a container and work within it, viewing and interacting with the project as if the machine they are accessing it from is identical to the machine specified in the container - regardless of what their computational environment actually is. They are designed to make it easier to transfer projects between very different environments.
In a way, containers behave like a virtual machine. To the outside world, they look like their own complete system. However, unlike a virtual machine, rather than creating a whole virtual operating system plus all the software and tools typically packaged with one, containers only contain the individual components they need in order to operate the project they contain. This gives a significant performance boost and reduces the size of the application.
Containers are a particularly useful way for reproducing research which relies on software to be configured in a certain way, or which makes use of libraries that vary between (or do not exist on) different systems. In summary, containers are a more robust way of sharing reproducible research than package management systems or Binder because they reproduce the entire system used for the research, not just the packages explicitly used by it. Their major downside is that due to their greater depth, they are conceptually more difficult to grasp and produce than many other methods of replicating computational environments.
Ben Corrie give a reasonably accessible overview of core concepts in ‘What is a container?’.
What are Images?#
Images are the files used to generate containers. Humans do not make images; they write recipes to generate images. Containers are then identical copies instantiated from images.
Think of it like this:
A recipe file a human writes contains all the steps to generate a working version of the project and its computational environment, but no actual materials. Think of this as a blueprint.
Building an image takes that recipe and using it, assembles all the packages, software libraries, and configurations needed to make the full-fledged project and environment, and bundles them up in a condensed lump. Think of images like a piece of flat-pack furniture made using the blueprint.
Containers take that image and assemble a fully working version of the project and the environment needed to run it. Think of this as assembling the flat-pack furniture.
So if a researcher wants to allow others to reproduce their work, they would need to write a recipe file, and use it to build an image of their project. They can then share this image file with anyone who wants to replicate their work. That person can then use the image to generate a container containing a working version of the project.
What is Docker?#
There are many tools available for creating and working with containers.
We will focus on Docker, which is widely used, but be aware that others such as Apptainer, LXC, Podman, Singularity also exist.
Apptainer and Singularity are designed with a focus on high-performance computing and tend to be well supported and preferred on such systems.
Podman may be seen as a completely free and open-source alternative to Docker.
It has a Docker compatible command-line interface and can run Docker container images.
Apptainer, Singularity and Podman do not need
sudo permissions to be run, while up until April 2020 Docker did (please see the Docker without root access section).
In Docker, the recipe files used to generate images are known as Dockerfiles, and should be named
Docker Hub hosts a great many pre-made images, such as images of Ubuntu machines, which can be downloaded and build upon. This makes the process of writing Dockerfiles relatively easy since users very rarely need to start from scratch, they can just customise existing images. However, this leaves a user vulnerable to similar security issues as described in the Security Issues of the YAML sub-chapter:
It is possible to include malicious code in Docker images
It is possible for people producing images to unknowingly include software in them with security vulnerabilities
This article goes deeper into the potential security vulnerabilities of containers and here is a detailed breakdown of security features currently within Docker, and how they function. The best advice for using images built by others is, as usual, only download and run something on your machine if it comes from a trusted source. Docker Hub has “official image” badges for commonly used, verified images as shown here:
Installers for Docker on a variety of different systems are available here. Detailed installation instructions are also available for a variety of operating systems such as Ubuntu, Debian, Macs, and Windows.
Here are a few key commands for creating and working with containers:
To build an image from a Dockerfile, go to the directory where the Dockerfile is and run:
sudo docker build --tag image_name .
To list the images on your system, use:
sudo docker image ls
To remove an image, run:
sudo docker rmi image_name
To open a container from an image, run:
sudo docker run -i -t image_name
-i -tflags automatically open up an interactive terminal within the container so you can view and interact with the project files.
To exit an interactive terminal, use:
To get a list of active containers with IDs, run:
sudo docker container ls
There are also three main commands used for changing the status of containers:
Pausing suspends the process running the container.
sudo docker pause container_ID
Containers can be unpaused by replacing
Stopping a container terminates the process running it. A container must be stopped before it can be deleted.
sudo docker stop container_ID
A stopped container can be restarted by replacing
stopdoes not work containers can be killed using
sudo docker kill container_ID
To remove a container, run:
sudo docker rm container_ID
Let us go through the anatomy of a very simple Dockerfile:
# Step 1: Set up the computational environment # Set the base image FROM ubuntu:18.04 # Install packages needed to run the project RUN apt-get update && \ apt-get install -y --no-install-recommends python3.7 python3-pip && \ rm -rf /var/lib/apt/lists/* RUN python3 -m pip install numpy #----------------------- # Step 2: Include the project files in the image # Copy files from the `project_files` directory on the machine building the image # into the `project` folder in the container. This folder and any missing # directories in its path are created automatically. COPY project_files/ project/
This looks complicated, but most of the lines in this example are comments (which are preceded by
There are only six lines of actual code.
The first of these is a
FROM statement specifying a base image.
All Dockerfiles require a FROM, even if it is just
All the following commands in a Dockerfile build upon the base image to make a functioning version of the researcher’s project.
Specifying a version for the image (
18.04 in this case) is optional.
However, it is best practice as it ensures that our Dockerfile remains valid after new releases of Ubuntu, which may not include packages (or specific versions thereof) that we require later (for example
It is worth spending time to choose an appropriate base image, as doing so can reduce the amount of work involved in writing a Dockerfile dramatically. For example, a collection of images with the R programming language included in them can be found here. If a project makes use of R, it is convenient to use one of these as a base image rather than spend time writing commands in your Dockerfile to install R.
The biggest block of lines comes next.
It’s a series of
RUN statements, which run shell commands when building the image.
In this block, they are used to install the software necessary to run the project. The first
RUN block is a
chain of commands of this form:
RUN command_to_do_thing_1 \ && command_to_do_thing_2 \ && command_to_do_thing_3 \ && command_to_do_thing_4
It is good practice to group related commands into a single
RUN block to reduce the final size of your image by
avoiding the creation of unnecessary layers.
We also follow best-practice by using
--no-install-recommends to avoid installing unnecessary packages
and cleaning up the
apt-cache, both of which further reduce the size of Debian or Ubuntu images.
After we have installed Python, we use another RUN statement to install a library required by our code.
COPY command is used to copy the project files from the machine building the image into the image itself.
The syntax of this command is
COPY file_to_copy location_in_container_to_copy_to.
In this example, all the files in the
project_files directory are included in the
project file in the container.
Note that you can only copy files from the directory where the Dockerfile is located, or subdirectories within it (in the example, that is the
ADD command has the same capabilities as
COPY, but it can also be used to add files not on the machine building the image.
For example it can be used to include files hosted online by following
ADD with a URL to the file.
It is good practice to use
COPY, except where
ADD is specifically required, as the term
COPY is more explicit about what is being done.
Here is what happens if a container is opened from an image called
book_example, built from the example above:
As you can see, the directory
project has been created, and inside the project files,
data.csv have been copied into it.
Because the Dockerfile already includes the software required for the project, in the image, the
analysis.py script runs without installing more software.
This command can be used in Dockerfiles to change the current working directory.
Commands that follow this in the Dockerfile will be applied within the new working directory unless/until another
WORKDIR changes the working directory.
When a container is opened with an interactive terminal, the terminal will open in the final working directory.
Here is a simple example of a Dockerfile that uses
WORKDIR, and the container it generates.
# Basic setup FROM ubuntu RUN apt-get update # Make a directory called A RUN mkdir A # Make the working directory A WORKDIR A # Make two directories, one called B_1 and one called B_2 RUN mkdir B_1 RUN mkdir B_2
B_2 have been created within directory
WORKDIR should be used when changing directories is necessary while building an image.
It may be tempting to use
RUN cd directory_name instead, as this syntax will be more familiar to those that commonly work via the command line,
but this can lead to errors.
RUN statement in a Dockerfile, the image is saved, and any following commands are applied to the image anew.
As an example, here is what happens in the above example if the
WORKDIR A line is swapped
RUN cd A.
All the directories have are in the top level in this case, rather than
B_2 being inside
This is because the image was restarted after the
RUN cd A command and opened at the top (root) level by default, so that is where the
mkdir B_1 and
mkdir B_2 commands took effect.
Other commands that are sometimes used in Dockerfiles include:
CMD: This is used to run commands as soon as the container is opened. This is different to RUN commands which are commands run as part of setting up a container. For example, to have a welcome message when a container is opened from the image,
CMDcould be used as follows:
CMD ["echo","Welcome! You just opened this container!"]
It is good practice to use CMD for any commands that need to be run before someone starts working in the container instead of forcing users to run them themselves (and trusting that they will even know that they need to).
VOLUMES: These will be discussed later.
MAINTAINER: This contains information regarding the person that wrote the Dockerfile. It is typically included at the top of a Dockerfile.
EXPOSE: This includes ports that should be exposed. It is more relevant to people using Docker to share web apps.
USER: Change the user that a command is run as (useful for dropping privileges).
Building Images and
As mentioned in the key commands section, to build an image open a terminal in the same directory as the Dockerfile to be used and run:
sudo docker build --tag name_to_give_image .
When an image is built everything in the Dockerfile’s directory and below (this is called the “context”) is sent to the Docker daemon to build the image.
The daemon uses the Dockerfile and its context to build the image.
If the context contains many large files, which are not needed for building the image, (old datafiles, for example) then it is a waste of time sending them to the daemon.
Doing so can make the process of building an image slow.
Files can be excluded from the context by listing them in a text file called
It is good practice to do so.
The files do not need to be listed individually in the
Here is an example of the contents of a
*.jpg **/*.png data_files/* file_to_exclude.txt
This excludes from the context:
.jpgfiles in the same directory as the Dockerfile file
.pngfiles in the same directory as the Dockerfile file or any subdirectories within it
All files within the
The file named
Copying Files To And From Containers#
Containers act much like virtual machines; as a result, copying files into and out of them is not as trivial as copying files to different locations within the same computer is.
A file can be copied from the machine running a container into the container using:
sudo docker cp file_name container_ID:path_to_where_to_put_file/file_name
Recall that container IDs can be obtained using
sudo docker container ls.
A file can be copied from within a container to the machine running the container by running the following command on the machine running the container:
sudo docker cp container_ID:path_to_file/file_name path_to_where_to_put_file/file_name
If the second part (the
path_to_where_to_put_file/file_name) is substituted for a
., then the file will be copied to whatever directory the terminal running the command is in.
Every time a container is opened from an image, that container is completely new. Say a container is opened, and work is done within it. If that container is closed, and the image it came from is again used to start another container, none of that work will be in the new one. It will simply have the starting state described in the image.
This can be a problem if a researcher wants to work in a container over time. Fortunately, there is a way around this using volumes. Volumes store the work done within a container even after it is closed, and can be used to load that work into future containers.
To create/use a volume, run:
sudo docker run -i -t --mount source=volume_name,target=/target_directory image_name
You should give your volume a more descriptive name than
target directory is required; only work within this directory will be saved in the volume.
Once the researcher is done, they can close the container as normal.
When they come back to the project and want to continue their work, they only need to use the same command as above, and it will load the work contained in
volume_name into the new container.
It will save any new work there too.
Below is a list of volume related commands:
To list volumes:
sudo docker volume ls
To delete a volume:
sudo docker volume rm volume_name
To delete all unattached volumes:
sudo docker volume prune
If, when deleting a container, a
-v is included after
sudo docker rm container_ID, any volumes associated with the container will also be deleted.
Docker without root access#
Up until April 2020, the only way to run Docker was with root access. “Rootless” mode was made available as part of the v20.10 release. Rootless mode is currently only available on Linux and requires an initial install of Docker >= v20.10.
The underlying difference between Docker without and with rootless mode is that previously any system running Docker had a daemon running as
uid0 that creates and owns all images, but with rootless mode the user creates and owns any images that they initialize.
To install and run the rootless version of Docker as a non-root user, use the following commands (where
20.10 refers to the installed version of Docker):
dockerd-rootless-setuptool.sh install docker run -d --name dind-rootless --privileged docker:20.10-dind-rootless
The following prequisites, which are part of the
shadow-utils package are required to run Docker rootless:
Podman is an open-source container engine that can be seen as a meaningful alternative to Docker.
In fact, Podman provides a Docker-compatible command-line interface.
For most use cases, a user familiar with Docker can simply alias the Docker command to Podman (
alias docker=podman) and carry on as normal.
Podman includes a full set of tools to create, run, manage, and share containers.
Podman is free and open-source software released under the Apache License 2.0. This is in contrast to Docker which has some open-source components, such as the engine and command-line interface, but also develops closed-source, subscription-requiring software, including the Docker Desktop clients for MacOS and Windows.
All Open Container Initiative container images can be used with Podman including Docker images hosted on Docker Hub. It is likely existing projects using Docker can be migrated to Podman.
Unlike Docker which uses a daemon running as root, Podman is daemonless. Unprivileged users can run containers using Podman. In most cases this will be configured automatically. This avoids a problem with the standard Docker configuration where users able to run containers have implicit access to the Docker daemon. The Docker Daemon is run by root by default and provides a trivial way to escalate privileges and get a high level of access to the hosts devices and filesystem.
Adding a user to the docker group is essentially giving them a high-level of access to the host with a very simple route to privilege escalation. For example:
$ docker run --mount=type=bind,source=/,destination=/host -it busybox
The user now has access to the filesystem of the host with the permissions of
The Podman documentation has up-to-date instructions for installing Podman.
It is important to understand that Podman is a tool for running Linux containers and so it requires a Linux host. If your computer is running Windows or MacOS, you can use the Podman remote client to interact with Podman on a Linux virtual machine or remote Linux Host.
Alternatively, the MacOS Podman client includes the experimental
podman machine subcommand for managing a Linux virtual machine that Podman can use.
Detailed instructions can be found on Podman’s GitHub repository
On Windows you can also run Podman in the Windows Subsystem for Linux (>= 2.0).
This RedHat blog post has instructions.
For most recent Linux distributions you should find Podman in the official repositories. Most distributions (including Arch, Debian, Fedora, and Ubuntu) will apply the appropriate configuration to let unprivileged users run Podman automatically. If there are any problems, the Podman documentation has instructions for configuration. This is as simple as two commands per user who should be able to run Podman.
Podman has a Docker-compatible command line interface so those commands will not be reiterated here.
The Docker commands in the key commands should all work by substituting
sudo docker with
Details of all commands and their options can be found in the Podman documentation.
Building Container Images#
Podman can build container images using the
podman build command.
Podman will build images from either Dockerfiles or Containerfiles.
Containerfiles use the same format as Dockerfiles, which are discussed in Writing Dockerfiles
Rootless containers are containers run by a normal user (not using
sudo or with the
These containers never have more privileges than the account that runs them.
This is an strong security advantage for rootless containers compared to running containers as root or through the Docker daemon.
If you are running a distribution with SELinux (for example Fedora or CentOS) you may need to add the
--privileged flag to the Podman commands below in order to access the host filesystem inside of containers.
This can be demonstrated with a simple example. Create a directory and put file with some text in it:
$ mkdir tmp $ echo "Hello" > tmp/a.txt
Now mount this directory into an interactive busybox container:
$ podman run --mount=type=bind,source=./tmp,destination=/tmp -it docker.io/library/busybox
In the container’s shell, confirm that the session belongs to the root user:
/ # id uid=0(root) gid=0(root) groups=10(wheel) / # whoami root
Append some text to the file created on the host, mounted at
/tmp/a.txt in the container:
/ # echo "World!" >> /tmp/a.txt
Create a new file in the
tmp directory and close the container:
/ # touch /tmp/b.txt / # exit
Inspect the files in
tmp on the host. The file
a.txt was modified by the container process:
$ cat tmp/a.txt Hello World!
b.txt was created. However, despite being created by
root inside the container, on the host it is owned by the user which ran the container.
This can be confirmed with
ls -l tmp/b.txt.
It is also impossible to read or modify files that the user running the container would not be able to.
For example, the
/etc/shadow file which contains users’ hashed passwords:
$ podman run --mount=type=bind,source=/etc/shadow,destination=/shadow -it docker.io/library/busybox / # cat /shadow cat: can't open '/shadow': Permission denied
If the above Podman command were run as root, using
sudo, then the container would not be rootless and it would be possible to read and modify
As Singularity is a tool for running Linux containers it can not run natively on Windows or MacOS.
Historically, a significant drawback of using Docker for reproducible research is that it was not intended as a user-space application but as a tool for server administrators. As such, it required root access to operate. There is, however, no reason why the execution of an analysis should require root access for the user. This is especially important when computations are conducted on a shared resource like high-performance computing (HPC) systems where users will never have root access.
The singularity container software was introduced to address this issue. Singularity was created with HPC systems and reproducible research in mind (see this video). Singularity containers do not require root access to run. Root access is normally required to build container images. However it is possible to build images as a normal user (with some restrictions) using the fakeroot feature. This enables users to build container images locally before running them on a high-performance cluster. As an added benefit, this makes it possible to bring software and project dependencies to an HPC system without requiring the system administrators to install or maintain them.
Furthermore, Singularity can take advantage of the large Docker ecosystem by building Singularity container images from Docker container images. Docker images can also be extend by building new images based on docker containers as a base layer.
Singularity Container Images#
Just as Docker images are built using
Dockerfile files, singularity containers are built from singularity definition files.
The Singularity documentation has a complete specification of the definition file format
As a minimal working example, we can build a
lolcow container based on the official ubuntu docker container image.
This is based on an example in the Singularity documentation.
Put the following text in a file named
Bootstrap: docker From: ubuntu %post apt-get -y update apt-get -y install fortune cowsay lolcat %environment export LC_ALL=C export PATH=/usr/games:$PATH %runscript fortune | cowsay | lolcat
This example uses a docker image (
ubuntu) as a basis.
The required packages are installed packages with
PATH variable is updated so that the run commands will be found when the container is run.
%runscript defines the command to be executed when the container is run.
A container image can then be built:
$ sudo singularity build lolcow.sif lolcow.def
This will pull the ubuntu image from Docker Hub, run the steps of the recipe in the definition file and produce a Singularity image file (
The container can be run with:
$ singularity run lolcow.sif
The way that Singularity packages container images as a single file in your working directory is convenient for migrating your work to HPC.
You may simply copy your container image to a cluster using
$ rsync -avz lolcow.sif <user>@<hpc_system>:~/
You should see a nice ASCII cow and a few words of wisdom:
___________________________________ / You will be called upon to help a \ \ friend in trouble. / ----------------------------------- \ ^__^ \ (oo)\_______ (__)\ )\/\ ||----w | || ||
A key distinction of Singularity is that it is able to natively utilise host GPUs in containers. Singularity has support for CUDA supporting GPUs from Nvidia and ROCm supporting GPUs from AMD. Running containers with GPU support does not require root privileges. Complete details on using GPUs can be found in the Singularity documentation
To use Nvidia GPUs in a container pass the
--nv flag to the
$ singularity pull docker://tensorflow/tensorflow:latest-gpu $ singularity exec --nv tensorflow_latest_gpu.sif nvidia-smi
Using AMD GPUs is similar but the
--rocm flag is used.
$ singularity pull docker://rocm/tensorflow:latest $ singularity run --rocm tensorflow_latest.sif
When using the
--rocm graphics drivers and libraries from the host are passed to the container.
You must therefore ensure that the application inside the container is compatible with the driver stack on the host.
For example, if the host has support for CUDA 10.2 a container featuring PyTorch build for CUDA 11.3 is likely to problems running.
Singularity on HPC#
Being HPC compatible, singularity containers are also supported by a wide range of workflow management tools. For example, both snakemake and nextflow support job-specific singularity containers. This makes singularity containers uniquely suited for parallelizing workflows on HPC systems using the widely used slurm workload manager. Using singularity, containers and snakemake/nextflow is a way of scaling reproducibility to a massive scale. Furthermore, as an added benefit, bringing workflows from a desktop machine to an HPC system no longer requires writing custom job submission scripts.
Long-term Storage of Container Images#
It is important to note that a mere container recipe file is not reproducible in itself since the build process depends on various (online) sources. Thus, the same recipe file might lead to different images if the underlying sources were updated.
To achieve true reproducibility, it isimportant to store the actual container images. For singularity images, this is particularly easy since an image is simply a large file. These can vary in size, from a few tens of megabytes (micro-containers) to several gigabytes, and are therefore not suited for being stored in a git repository themselves A free, citable, and long-term solution to storing container images is zenodo.org which allows up to 50 Gb per repository. Since zenodo mints DOIs for all content uploaded, the images are immediately citable. In contrast to Docker Hub (which also only accepts docker images), zenodo is also clearly geared towards long-term storage and discoverability via a sophisticated metadata system. Thus, it is ideally suited for storing scientific containers associated with particular analyses since these tend to not change over time.
Words of Warning#
Even though singularity and docker might look similar, they are conceptually very different. Singularity handles the distinction between the host and container file system differently. For instance, by default, singularity includes a few bind points in the container, namely:
$PWD comes in handy since it implies that all files in the working directory are visible within the container.
$HOME by default, however, also implies that software using configuration files from
$HOME might behave unexpectedly since the image specific configuration files are overwritten with the current users settings in
While this behaviour is handy in HPC scenarios, it is potentially dangerous for reproducible research.
To avoid potential issues, any software installed in a singularity container should be pointed to a global, user-independent configuration file.