Using Singularity Containers
This document describes how to use Singularity, a software tool that allows you to run Linux containers. Using containers facilitates the movement of software applications and workflows between various computational environments.
Singularity is a software tool provided on the Savio cluster. It allows you to bring already-built research applications and workflows from other Linux environments onto Savio and run them on the cluster, without any installation or reconfiguration required. Singularity packages those applications and workflows in “containers,” and runs them within the Singularity container's boundaries.
- Containerization provides "lightweight, standalone, executable packages of software that include everything needed to run an application: code, runtime, system tools, system libraries and settings".
- A container provides a self-contained (isolated) filesystem.
- Containers are similar to virtual machines in some ways, but much lighter-weight.
- Containers are portable, shareable, and reproducible.
Singularity allows you to create containers, or find and obtain containers from others, and then run them on any Linux platform where Singularity is installed. Research software that you or others have packaged up into Singularity containers can be copied to -- and run on -- multiple clusters, cloud environments, workstations, and laptops.
Singularity thus enables “Bring Your Own Environment” computing. It is conceptually similar to Docker, a well-known software containerization platform that isn’t compatible with the security models used on Savio and other traditional High Performance Computing (HPC) environments. Both Singularity and Docker, in turn, have some similarities to virtual machines.
Singularity containers that you use on Savio must be created on a different computer. Root permission is required to create Singularity containers, and users are not allowed to run as root on the cluster. Options for creating image-based Singularity containers, which can then be run on Savio under a user’s normal set of permissions, are described below. One option includes using existing Docker images directly in Singularity on Savio.
In addition to this documentation, more information can be found in our April 2021 training on using Singularity on Savio.
Running Singularity containers on Savio¶
Trying things out on a login node¶
Assuming you have a Singularity container in a directory on Savio you can run it as follows.
singularity run mycontainer.sif
We can run a Docker container available from DockerHub (behind the scenes the Docker image will be downloaded and converted to a Singularity image) like this:
singularity run docker://ubuntu:20.04
That will put you into a shell inside a container running Ubuntu Linux 20.04. Note the change in prompt after the container starts. Inside the container you could do things like the following to convince yourself that you are running in the container and not on Savio, although your working directory will generally be a Savio directory.
cat /etc/issue # not the Savio OS! which python # not much here! pwd
Singularity containers can be used in three ways:
- shell sub-command: invokes an interactive shell within a container
singularity shell mycontainer.sif
- run sub-command: executes the container’s runscript (i.e., the primary way the container's builder intends for the container to be used)
singularity run mycontainer.sif
- exec sub-command: execute an arbitrary command within container
singularity exec mycontainer.sif cat /etc/os-release
In the example with ubuntu:20.04 above, the container's runscript simply starts a shell inside the container.
Running via Slurm¶
Of course in most cases you will be running jobs under Slurm.
To submit a batch job that runs Singularity on Savio, create a SLURM job script file. A simple example follows:
# Job name: #SBATCH --job-name=test_singularity # # Account: #SBATCH --account=account_name # # Partition: #SBATCH --partition=partition_name # # Wall clock limit: #SBATCH --time=00:30:00 # ## Command(s) to run: singularity run /path/to/container/mycontainer.sif
Creating Singularity container images¶
You have a variety of options for creating Singularity container images that you can run on Savio.
- Directly create a Singularity image by importing a Docker or Singularity image from an image registry.
- Docker image registries include DockerHub and the GitHub container registry
- Singularity image registries include SingularityHub or the Sylabs container registry. Docker image registry such as DockerHub.
- This does not require root access in any form but restricts you to only the images already available.
- Create a Docker image on your own machine.
- One option is then to push it to a Docker registry (such as DockerHub), and then import the Docker image as above.
- A second option is to archive the Docker image, transfer it to Savio, and then convert to a Singularity image.
- These options rely on installing and running Docker on your own machine.
- Create a Singularity image on your own machine and transfer it to Savio.
- One option is to install Singularity on your own machine.
- A second option is to install Docker on your own machine and run Singularity within the quay.io Singularity Docker image.
- Use a cloud service that allows you to build images, such as Sylabs Remote Builder
Here we provide more details on some of these options. More details can be found in our training and in various online documentation for Docker and Singularity.
Note that when building an image from either a Dockerfile or Singularity definition file, you generally want to base your image (i.e., bootstrap it) on an existing image that may have key software already installed. Some examples include images with Tensorflow, PyTorch or R/RStudio.
Import an existing Docker image or Singularity image¶
- You can simply ask Singularity to run a Docker container as shown above; behind the scenes this will create a Singularity image file.
To create the Singularity image and run in two explicit steps:
singularity pull docker://ubuntu:20.04 # this creates ubuntu_20.04.sif singularity run ubuntu_20.04.sif
Note that this also creates a copy of the container in your Singularity cache, which can quickly fill up your home directory. Some useful comands for working with the cache are:
singularity cache list singularity cache clean
You can control where Singularity locates the cache and the temporary directory that Singularity uses by setting
SINGULARITY_TMPDIR, respectively. In particular, you may want to use your scratch directory or
/tmp as the location of the cache.
- Here's how to run an existing Singularity image from a registry:
singularity pull hello-world.sif shub://singularityhub/hello-world singularity run hello-world.sif
Create a Docker image on your own machine¶
For this you will need a Dockerfile that defines what you want in your image. There are lots of examples online.
- To create an image and push to a Docker image registry (in this case DockerHub), the process looks like this, assuming you have a Dockerfile in your working directory:
docker login --username=paciorek docker build -t tagname . docker tag tagname paciorek/name_of_image:0.1 # version number is set to be 0.1 docker push paciorek/name_of_image
This assumes you have created a DockerHub account; here the DockerHub username is
Once the image is on the registry you can run it (or pull and then run it) using the commands shown above for pre-existing Docker images.
- To create an image, archive it locally on the machine you are using, and convert to a Singularity image, the process looks like this:
docker build -t tagname . docker save tagname > name_of_image.tar
Then transfer the .tar file to Savio and run:
singularity build name_of_image.sif docker-archive://name_of_image.tar
Create a Singularity image on your own machine¶
For this you will need to create a Singularity definition file. Please see the Singularity documentation for more details about these or our training for an example.
If you are running Singularity directly on a machine where you have root access, you can build from the definition file like this:
singularity build alpine-example.sif alpine-example.def
If you are using the Singularity Docker image, it would look like this:
docker run --privileged -t --rm -v $PWD:/app quay.io/singularity/singularity:v3.7.1 \ build /app/alpine-example.sif /app/alpine-example.def
In either case, you then simply transfer the resulting .sif image file to Savio.
Accessing your Savio storage from within Singularity containers¶
It can be useful for scripts running within Singularity to reference directories outside the Singularity container, i.e., directories on the Savio filesystem. In fact, when using a container you would generally do input and output to/from files on the Savio filesystem rather than files in the container.
- A user's home directory and scratch directory (as well as
/tmp) are automatically available inside the container via the usual paths,
- Your working directory in the container will generally be the working directory on Savio from which you started the container (or in some cases simply your Savio home directory).
To reference other directories from within the container you need to create your own 'bind paths' that indicate which directory on the Savio filesystem to associate with a path in the container file system. The basic syntax is:
For example here we start a shell inside a container that mounts a subdirectory of a user's Savio scratch directory to
/data in the container and creates a new file called
erase-me that is accessible at
/global/scratch/paciorek/some_dir/erase-me outside the container.
singularity shell -B /global/scratch/paciorek/some_dir:/data hello-world.sif ls /data echo "hello from inside the container" >> /data/erase-me exit ls -l /global/scratch/paciorek/some_dir/erase-me
Using MPI with Singularity¶
You can run Singularity containers via MPI. You'll need to have MPI installed within the container.
- If you are working on a single node, you can run MPI within a container.
- However, more commonly you would use the MPI executable on Savio to execute Singularity containers.
The key thing in order to use the system MPI to run Singularity containers is to make sure the MPI installed inside the container is compatible with the MPI installed on Savio. The easiest way to ensure this is to have the version inside the container be the same version as the MPI module you plan to use on Savio. You can see these modules with:
module load gcc # load the gcc version of interest module avail openmpi # see the MPI versions available for that gcc
Here is an example of running a Singularity container via MPI:
module load gcc openmpi mpirun singularity exec my_singularity_container_with_mpi.sif \ /path/to/my/mpi/executable
That will launch
/path/to/my/mpi/executable (which should be on Savio, not in the container) on as many processes as the number of tasks specified in your Slurm job.
Using Singularity with GPUs¶
You can easily use a Singularity container that does computation on a GPU.
Singularity supports NVIDIA’s CUDA GPU compute framework or AMD’s ROCm solution.
By using the
--nv flag when running Singularity, the NVIDIA drivers on Savio are dynamically mounted into the container at run time. The container should provide the CUDA toolkit, using a version of the toolkit that is compatible with the NVIDIA driver version on Savio.
The minimal driver requirement for a specific version of the CUDA runtime/toolkit can be found in Table 1 here. E.g., CUDA 11.2 requires NVIDIA driver version >= 450.80.02.
Savio's NVIDIA driver version can be found by running
nvidia-smi on a GPU node. Currently Savio has version 440.44, so one would not want to use or create a container with CUDA 11.2, but one could use CUDA 10.2
Here's an example of running a Singularity container based on a Docker container that provides GPU-using software. I am using an older version of PyTorch because newer versions depend on CUDA versions not supported by Savio's NVIDIA driver version.
singularity run --nv docker://pytorch/pytorch:1.6.0-cuda10.1-cudnn7-runtime
Of course it only makes sense to do this after using
srun to get access to a node with a GPU.