User manual

A basic user tool to execute simple Docker containers in user space without requiring root privileges. udocker enables basic download and execution of Docker containers by non-privileged users in Linux systems where Docker is not available. It can be used to access and execute the content of docker containers in Linux batch systems and interactive clusters that are managed by other entities such as grid infrastructures, HPC clusters or other externally managed batch or interactive systems.

udocker does not require any type of privileges nor the deployment of services by system administrators. It can be downloaded and executed entirely by the end user. The limited root functionality provided by some of the udocker execution modes is either simulated or provided via user namespaces.

udocker is a wrapper around several tools and technologies to mimic a subset of the Docker capabilities including pulling images and running then with minimal functionality.

udocker is mainly meant to execute user applications packaged in Docker containers. We recommend the use of Docker whenever possible, but when it is unavailable udocker can be the right tool to run your applications.

1. Introduction

1.1. How does it work

udocker is written in Python, since v1.3.0 (or the development v1.2.x), udocker supports Python 2.6, 2.7 and Python >= 3.6. udocker has a minimal set of dependencies so that can be executed in a wide range of Linux systems. udocker does not make use of Docker nor requires its installation.

udocker “executes” the containers by simply providing a chroot like environment to the extracted container. udocker is meant to integrate several technologies and approaches hence providing an integrated environment that offers several execution options. This version provides execution engines based on PRoot, Fakechroot, runc, crun and Singularity to facilitate the execution of Docker containers without privileges.

The basic usage flow starts by downloading the image from an image repository in the usual way; create the container out of that image (flattening the image on the filesystem), and finally run the container with the name we gave it in the creation process:

This sequence allows the created container to be executed many times. If simultaneous executions are envisage just make sure that input/output files are not overwritten by giving them different names during execution as the container will be shared among executions.

Containers can also be pulled, created and executed in a single step. However in this case a new container is created for every run invocation thus occupying more storage space. To pull, create and execute in a single step invoke run with an image name instead of container name:

1.2. Limitations

Since root privileges are not involved, any operation that really requires privileges is not possible. The following are examples of operations that are not possible:

Other limitations:

1.3. Security

udocker does not offer robust isolation features such as the ones offered by docker. Therefore if the containers content is not trusted then these containers should not be executed with udocker as they will run inside the user environment. For this reason udocker should not be run by privileged users.

udocker does not require privileges and runs under the identity of the user invoking it.

The containers data will be unpacked and stored in the user home directory or other location of choice. Therefore the containers data will be subjected to the same filesystem protections as other files owned by the user. If the containers have sensitive information the files and directories should be adequately protected by the user.

Users can download the udocker tarball, install in the home directory and execute it from their own accounts without requiring system administration intervention.

udocker provides a chroot like environment for container execution. This is currently implemented by:

udocker via PRoot offers the emulation of the root user. This emulation mimics a real root user (e.g getuid will return 0). This is just an emulation no root privileges are involved. This feature enables tools that do not require privileges but that check the user id to work properly. This enables for instance software installation with rpm and yum inside the container.

Similarly to Docker, the login credentials for private repositories are stored in a file and can be easily accessed. Logout can be used to delete the credentials. If the host system is not trustable the login feature should not be used as it may expose the login credentials.

udocker does not have privileged escalation issues as it runs entirely without privileges.

1.4. Basic flow

The basic flow with udocker is:

  1. The user downloads udocker to its home directory and executes it
  2. Upon the first execution udocker will download additional tools
  3. Container images can be fetched from Docker Hub with pull
  4. Containers can be created from the images with create
  5. Containers can then be executed with run

Additionally:

2. Installation

udocker can be deployed in the user home directory and thus does not require system installation. For further information see the Installation manual.

3. Commands

3.1. Syntax

The udocker syntax is very similar to Docker. Since version 1.0.1 the udocker preferred command name changed from udocker.py to udocker. A symbolic link between udocker and maincmd.py is provided when installing with the distribution tarball.

udocker [GLOBAL-PARAMETERS] COMMAND [COMMAND-OPTIONS] [COMMAND-ARGUMENTS]

Quick examples:

udocker --help
udocker run --help

udocker pull busybox
udocker --insecure pull busybox
udocker create --name=verybusy busybox
udocker run -v /tmp/mydir verybusy
udocker run verybusy /bin/ls -l /etc

udocker pull --registry=https://registry.access.redhat.com  rhel7
udocker create --name=rh7 rhel7
udocker run rh7

3.2. Obtaining help

General help about available commands can be obtained with:

udocker --help

Command specific help can be obtained with:

udocker COMMAND --help

3.3. install

udocker install [OPTIONS]

Install of udocker tools. Pulls the tools and installs them in the user home directory under $HOME/.udocker or in a location defined by the environment variable UDOCKER_DIR. The pulling may attempt several mirrors.

Options:

Examples:

udocker install
udocker install --force --purge

udocker search [-a] STRING
udocker search --list-tags REPO/IMAGE

Search Docker Hub for container images. The command displays containers one page at a time and pauses for user input. Not all registries have search capabilities.

Options:

Examples:

udocker search busybox
udocker search -a busybox
udocker search iscampos/openqcd
udocker search --list-tags centos

3.5. pull

udocker pull [OPTIONS] REPO/IMAGE:TAG

Pull a container image from a docker repository by default uses dockerhub. The associated layers and metadata are downloaded from dockerhub. Requires python pycurl or the presence of the curl command.

Options:

Examples:

udocker pull busybox
udocker pull fedora:latest
udocker pull indigodatacloudapps/disvis
udocker pull quay.io/something/somewhere
udocker pull --httpproxy=socks4://host:port busybox
udocker pull --httpproxy=socks5://host:port busybox
udocker pull --httpproxy=socks4://user:pass@host:port busybox
udocker pull --httpproxy=socks5://user:pass@host:port busybox
udocker pull --httpproxy=socks4a://host:port busybox
udocker pull --httpproxy=socks5h://host:port busybox
udocker pull --httpproxy=socks4a://user:pass@host:port busybox
udocker pull --httpproxy=socks5h://user:pass@host:port busybox
udocker pull --platform=linux/arm64 fedora:latest
udocker pull --platform=linux/ppc64le centos:7

3.6. images

udocker images [OPTIONS]

List images available in the local repository, these are images pulled form Docker Hub, and/or load or imported from files.

Options:

Examples:

udocker images
udocker images -l

3.7. create

udocker create [OPTIONS] REPO/IMAGE:TAG

Extract a container from an image available in the local repository. Requires that the image has been previously pulled from Docker Hub, and/or load or imported into the local repository from a file. use udocker images to see the images available to create. If successful the command prints the id of the extracted container. An easier to remember name can also be given with --name.

Options:

Examples:

udocker create --name=mycontainer indigodatacloud/disvis:latest

3.8. ps

udocker ps [options]

List extracted containers. These are not processes but containers extracted and available to the executed with udocker run. The command displays:

Options:

Examples:

udocker ps

3.9. rmi

udocker rmi [OPTIONS] REPO/IMAGE:TAG

Delete a local container image previously pulled/loaded/imported. Existing images in the local repository can be listed with udocker images. If short of disk space deleting the image after creating the container can be an option.

Options:

Examples:

udocker rmi -f indigodatacloud/ambertools\_app:latest

3.10. rm

udocker rm [options] CONTAINER-ID

Delete a previously created container. Removes the entire directory tree extracted from the container image and associated metadata. The data in the container tree WILL BE LOST. The container id or name can be used.

Options:

Examples:

udocker rm 7b2d4456-9ee7-3138-ad01-63d1342d8545
udocker rm mycontainer

3.11. inspect

udocker inspect REPO/IMAGE:TAG
udocker inspect [OPTIONS] CONTAINER-ID

Prints container metadata. Applies both to container images or to previously extracted containers, accepts both an image or container id as input.

Options:

Examples:

udocker inspect ubuntu:latest
udocker inspect d2578feb-acfc-37e0-8561-47335f85e46d
udocker inspect -p d2578feb-acfc-37e0-8561-47335f85e46d

3.12. name

udocker name CONTAINER-ID NAME

Give an easier to remember name to an extracted container. This is an alternative to the use of create --name=

Examples:

udocker name d2578feb-acfc-37e0-8561-47335f85e46d BLUE

3.13. rmname

udocker rmname NAME

Remove a name previously given to an extracted container with udocker --name= or with udocker name. Does not remove the container.

Examples:

udocker rmname BLUE

3.14. rename

udocker rename NAME NEWNAME

Change a container name previously given to an extracted container with udocker --name= or with udocker name. Does not change the container id.

Examples:

udocker rename BLUE GREEN

3.15. verify

udocker verify REPO/IMAGE:TAG

Performs sanity checks to verify a image available in the local repository.

Examples:

udocker verify indigodatacloud/powerfit:latest

3.16. import

udocker import [OPTIONS] TARBALL|- REPO/IMAGE:TAG

Import a tarball from file or stdin. The tarball can be imported into a new image or container. Without options can be used to import a container exported by Docker (with docker export) creating a new image in the local repository. When using --tocontainer allows importing directly into containers without creating images in the local repository. Use --tocontainer alone to import a container exported by docker (with docker export) into a new container without creating an image. Use --clone to import a udocker container (e.g. exported with udocker export --clone) into a new container also without creating an image and allowing to preserve the container metadata and udocker execution modes. The option --name= adds a name alias to the created container, is used in conjunction with --tocontainer or --clone.

Options:

Examples:

udocker import docker_container.tar myrepo:latest
udocker import - myrepo:latest < docker_container.tar
udocker import --mv docker_container.tar myrepo:latest
udocker import --tocontainer --name=BLUE docker_container.tar
udocker import --clone --name=RED udocker_container.tar

3.17. load

udocker load -i IMAGE-FILE
udocker load -i IMAGE-FILE NAME
udocker load -

Loads into the local repository a tarball containing a Docker image with its layers and metadata. This is equivalent to pulling an image from Docker Hub but instead loading from a locally available file. It can be used to load a Docker image saved with docker save. A typical saved image is a tarball containing additional tar files corresponding to the layers and metadata. From version 1.1.4 onwards, udocker can also load images in OCI format. The optional NAME argument can be used to change the name of the loaded image. This argument is particularly relevant to provide adequate names to OCI loaded images as these frequently only provide tag names. If an OCI image does not provide a name and the argument NAME is also not provided in the command line, then udocker will generate a random name.

Examples:

udocker load -i docker-image.tar
udocker load - < docker-image.tar
udocker load -i oci-image.tar test-image

3.18. protect

udocker protect REPO/IMAGE:TAG
udocker protect CONTAINER-ID

Marks an image or container against deletion by udocker. Prevents udocker rmi and udocker rm from removing images or containers.

Examples:

udocker protect indigodatacloud/ambertools\_app:latest
udocker protect 3d528987-a51e-331a-94a0-d278bacf79d9

3.19. unprotect

udocker unprotect REPO/IMAGE:TAG
udocker unprotect CONTAINER-ID

Removes a mark against deletion placed by udocker protect.

Examples:

udocker unprotect indigodatacloud/ambertools\_app:latest
udocker unprotect 3d528987-a51e-331a-94a0-d278bacf79d9

3.20. mkrepo

udocker mkrepo DIRECTORY

Creates a udocker local repository in specify directory other than the default one ($HOME/.udocker). Can be used to place the containers in another filesystem. The created repository can then be accessed with udocker --repo=DIRECTORY COMMAND.

Examples:

udocker mkrepo /tmp/myrepo
udocker --repo=/tmp/myrepo pull docker.io/fedora/memcached
udocker --repo=/tmp/myrepo images

3.21. run

udocker run [OPTIONS] CONTAINER-ID|CONTAINER-NAME
udocker run [OPTIONS] REPO/IMAGE:TAG

Executes a container. The execution several execution engines are provided. The container can be specified using the container id or its associated name. Additionally it is possible to invoke run with an image name, in this case the image is extracted and run is invoked over the newly extracted container. Using this later approach will create multiple container directory trees possibly occupying considerable disk space, therefore the recommended approach is to first extract a container using udocker create and only then execute with udocker run. The same extracted container can then be executed as many times as required without duplication.

udocker provides several execution modes to support the actual execution within a container. Execution modes can be changed using the command udocker setup --execmode=<mode> <container-id> for more information on available modes and their characteristics see section 3.27.

Options:

Options valid only in Pn execution modes:

Options valid only in Rn execution modes:

Examples:

# Pull fedora from Docker Hub
udocker pull fedora:29

# create the container named myfed from the image named fedora
udocker create --name=myfed  fedora:29

# execute a cat inside of the container
udocker run  myfed  cat /etc/redhat-release

# The above three operations can be done with a single command
# However each time udocker is invoked in this way a new container
# directory tree is created. This will consume additional space
# and may considerably increase the time for the container to start.
udocker run fedora:29 cat /etc/redhat-release

# For repeated invocations of the same container image the issue
# described above can be prevented by using --pull=reuse with --name.
# With the option --pull=reuse udocker will first try to execute
# a container with the same name specified by --name and only if
# it doesn't exist will it pull and create. In this way repeated
# calls to run only create a single container that is then reused.
udocker run --name=F29 --pull=reuse fedora:29 cat /etc/redhat-release

# In this example the host /tmp is mapped to the container /tmp
udocker run --volume=/tmp  myfed  /bin/bash

# Same as above but running something in /tmp
udocker run  -v=/tmp  myfed  /bin/bash -c "cd /tmp; ./myscript.sh"

# Run binding a host directory inside the container to make it available
# The host $HOME is mapped to /home/user inside the container
# The shortest -v form is used instead of --volume=
# The option -w same as --workdir is used to change dir to /home/user
udocker run -v=$HOME:/home/user -w=/home/user myfed  /bin/bash

# Install software inside the container
udocker run  --user=root myfed  yum install -y firefox pulseaudio gnash-plugin

# Run as certain uid:gid inside the container
udocker run --user=1000:1001  myfed  /bin/id

# Run firefox
udocker run --bindhome --hostauth --hostenv \
    -v /sys -v /proc -v /var/run -v /dev --user=green --dri myfed  firefox

# Run in a script
udocker run ubuntu  /bin/bash <<EOF
cd /etc
cat motd
cat lsb-release
EOF

# Search and pull from another repository than dockerhub
# First search for the expression `myrepo` in quay.io
# Second list the tags for a given image in quay.io
# Third finally pull a given image:tag from quay.io
udocker search quay.io/myrepo
udocker search --list-tags quay.io/myrepository/myimage
udocker pull quay.io/myrepository/myimage:v2.3.1

# Run container in a given directory tree using the DEFAULT EXECUTION MODE
# Below ROOT is the complete directory structure of the container operating system
# This enables udocker to execute directory trees created by other tools
# Much of the udocker functionality is not usable when using --location
./udocker run --location=/tmp/u/containers/07b3226e-6513-3f85-884f-e3cfdd2fbc0e/ROOT

3.22. Debug and Verbosity

Further debugging information can be obtaining by running with -D.

Examples:

udocker -D pull busybox:latest
udocker -D run busybox:latest

The options -q or --quiet can be specified before each command to reduce verbosity. The verbosity level can also be specified by assigning a value between 0 and 5 to the environment variable UDOCKER_LOGLEVEL.

Examples:

udocker -q run busybox:latest /bin/ls
UDOCKER_LOGLEVEL=2 udocker run busybox:latest /bin/ls

3.23. login

udocker login [--username=USERNAME] [--password=PASSWORD | --password-stdin ] [--registry=REGISTRY]

Login into a Docker registry using v2 API. Only basic authentication using username and password is supported. The username and password can be prompted or specified in the command line. The username is the username in the repository, not the associated email address.

Options:

Examples:

# To use dockerhub private repositories
udocker login --username=xxxx --password=yyyy

# To use a different container registry (the https:// is optional)
udocker login --registry=https://hostname
username: xxxx
password: ****

# To use a private repository at AWS ECR
aws ecr get-login-password --region eu-north-1 | udocker login --username=AWS --password-stdin --registry=000000000000.dkr.ecr.eu-north-1.amazonaws.com

3.24. logout

udocker logout [-a]

Delete the login credentials (username and password) stored by previous logins. Without arguments deletes the credentials for the current registry. To delete all registry credentials use -a.

Options:

Examples:

udocker logout
udocker logout --registry="https://hostname:5000"
udocker logout -a

3.25. clone

udocker clone [--name=NAME] CONTAINER-ID|CONTAINER-NAME

Duplicate an existing container creating a complete replica. The replica receives a different CONTAINER-ID. An alias can be assigned to the newly created container by using --name=NAME.

Options:

Examples:

udocker clone f24771be-f0bb-3046-80f0-db301e099517
udocker clone --name=RED  f24771be-f0bb-3046-80f0-db301e099517
udocker clone --name=RED  BLUE

3.26. save

udocker save -o IMAGE-FILE REPO/IMAGE:TAG
udocker save -o - REPO/IMAGE:TAG

Saves an image including all its layers and metadata to a tarball. The input is an image not a container, to produce a tarball of a container use export. The saved images can be read by udocker or Docker using the command load.

Examples:

udocker save -o docker-image.tar centos:centos7
udocker save -o - > docker-image.tar ubuntu:16.04 ubuntu:18.04 ubuntu:19.04

3.27. setup

udocker setup [--execmode=XY] [--force] [--nvidia] [--purge] CONTAINER-ID|CONTAINER-NAME

With --execmode chooses an execution mode to define how a given container will be executed, namely enables selection of an execution engine and its related execution modes. Without options, setup will print the current execution mode for the given container. The option --nvidia enables access to GPGPUs by adding the necessary host libraries to the container. The option --force can be used both with --execmode and with --nvidia to force the setup of the container to the specified mode. The option --purge removes mount points, auxiliary files and directories created by udocker inside the container directory tree to support its execution. It should only be invoked when there is no execution taking place as it may affect processes running in the container tree.

Options:

Mode Engine Description Changes container
P1 PRoot accelerated mode using seccomp No
P2 PRoot seccomp accelerated mode disabled No
F1 Fakechroot exec with direct loader invocation symbolic links
F2 Fakechroot F1 plus modified loader F1 + ld.so
F3 Fakechroot fix ELF headers in binaries F2 + ELF headers
F4 Fakechroot F3 plus enables new executables and libs same as F3
R1 runc rootless user mode namespaces resolv, passwd
R2 runc R1 plus P1 for software installation resolv, passwd, proot
R3 runc R1 plus P2 for software installation resolv, passwd, proot
S1 Singularity uses singularity if available in the host passwd

The default execution mode is P1 using PRoot and starting in root emulation mode.

The mode P2 also uses PRoot and although has lower performance than P1 can be more reliable. The mode P1 uses PRoot with SECCOMP syscall filtering which provides higher performance in most operating systems. PRoot provides the most universal execution mode in udocker but may also exhibit lower performance on older kernels such as in CentOS 6 systems. The Pn modes also offer root emulation to facilitate software installation and to execute applications that expect to run under root.

The Fakechroot (Fn), runC (Rn) and Singularity (Sn) engines are EXPERIMENTAL. They provide higher performance in most cases, but are less universal thus supporting less Linux distributions.

The udocker Fakechroot engine has four modes that offer increasing compatibility levels. F1 is the least intrusive mode and only changes absolute symbolic links so that they point to locations inside the container. F2 adds changes to the loader to prevent loading of host shareable libraries. F3 adds changes to all binaries (ELF headers of executables and libraries) to remove absolute references pointing to the host shareable libraries. These changes are performed once during the setup, executables added after setup will not have their ELF headers fixed and will fail to run. Notice that setup can be rerun with the --force option to fix these binaries. F4 performs the ELF header changes dynamically (on-the-fly) thus enabling compilation and linking within the container and new executables to be transferred to the container and executed. Executables and libraries in host volumes are not changed and hence cannot be executed from a container in F2, F3 and F4 execution modes. runC with rootless user namespaces requires a recent Linux kernel and is known to work on Ubuntu and Fedora hosts.

Mode Rn requires kernels with support for rootless containers, thus it will not work on some distributions (e.g. CentOS 6 and CentOS 7). The rootless execution modes have inherent limitations related to the manipulation of uids and gids that may cause certain operations to fail such as software installations. To overcome this limitation of the R1 execution mode, udocker provides the R2 and R3 execution modes that combine runc with the proot uid/gid emulation. In these modes the execution chain is:

runc -> proot -> executable

When using the Rn modes, udocker will search for a runc executable in the host system, only if it does not find one it will default to use the runc provided with the udocker tools. This behavior can be change through environment variables and configuration settings. Fakechroot requires libraries compiled for each guest operating system, udocker provides these libraries for several distributions including Ubuntu 14, Ubuntu 16, Ubuntu 18, CentOS 6 and CentOS 7 and some others. Other guests may or may not work with these same libraries.

Notice that changes performed in Fn and Rn modes will prevent the containers from running in hosts where the directory path to the container is different. In this case convert back to P1 or P2, transfer to the target host, and then convert again from Pn to the desired Fn mode.

Singularity must be available in the host system for execution mode S1. Newer versions of Singularity may run without requiring privileges but need a recent kernel in the host system with support for rootless user mode namespaces similar to runc in mode R1. Singularity cannot be compiled statically due to dependencies on dynamic libraries and therefore is not shipped with udocker. In CentOS 6 and CentOS 7 Singularity must be installed with privileges by a system administrator as it requires suid or capabilities. The S1 mode also offers root emulation to facilitate software installation and to execute applications that expected to run under root.

Examples:

udocker create --name=mycontainer  fedora:25

udocker setup --execmode=F3  mycontainer
udocker setup  mycontainer                 # prints the execution mode

udocker run  mycontainer /bin/ls

udocker setup  --execmode=F4  mycontainer
udocker run  mycontainer /bin/ls

udocker setup  --execmode=P1  mycontainer
udocker run  mycontainer  /bin/ls

udocker setup  --execmode=R1  mycontainer
udocker run  mycontainer  /bin/ls

udocker setup  --execmode=S1  mycontainer
udocker run  mycontainer  /bin/ls

The default execution mode of udocker can also be changed. This has however several limitations, therefore the recommended method to change the execution mode is via the udocker setup command. The default execution mode can be changed through the configuration files by changing the attribute default_execution_mode or through the environment variable UDOCKER_DEFAULT_EXECUTION_MODE. Only the following modes can be used as default modes: P1, P2, F1, S1, and R1. Changing the default execution mode can be useful if the default does not work as expected.

Example:

UDOCKER_DEFAULT_EXECUTION_MODE=P2 ./udocker run mycontainer /bin/ls

3.28. tag

udocker tag SOURCEREPO/IMAGE:TAG  TARGETREPO/IMAGE:TAG

Creates a new image tag from an existing source image. The newly created image tag is a replica of the source image. The source image can be removed or further updated via pull without affecting the newly created tag. A new tag does not occupy additional space as the image layers are shared. The image layers are only removed from the local udocker repository when no other image is referencing them.

Example:

udocker tag centos:centos7  mycentos:mycentos7

3.29. manifest inspect

udocker manifest inspect REPO/IMAGE:TAG

Obtain and print information about an IMAGE manifest from a remote registry. Can be used to obtain the platform architectures supported by the IMAGE.

Options:

Example:

udocker manifest inspect centos:centos7
udocker manifest --platform=linux/ppc64le inspect centos:7

4. Running MPI jobs

In this section we will use the Lattice QCD simulation software openQCD to demonstrate how to run Open MPI applications with udocker (http://luscher.web.cern.ch/luscher/openQCD). Lattice QCD simulations are performed on high-performance parallel computers with hundreds and thousands of processing units. All the software environment that is needed for openQCD is a compliant C compiler and a local MPI installation such as Open MPI.

In what follows we describe the steps to execute openQCD using udocker in a HPC system with a batch system (e.g. SLURM). An analogous procedure can be followed for other MPI applications.

A container image of openQCD can be downloaded from the Docker Hub repository. From this image a container can be extracted to the filesystem (using udocker create) as described below.

./udocker pull iscampos/openqcd
./udocker create --name=openqcd iscampos/openqcd
fbeb130b-9f14-3a9d-9962-089b4acf3ea8

Next the created container is executed (notice that the variable LD_LIBRARY_PATH is explicitly set):

./udocker run -e LD_LIBRARY_PATH=/usr/lib openqcd /bin/bash

In this approach the host mpiexec will submit the N MPI process instances, as containers, in such a way that the containers are able to communicate via the low latency interconnect (Infiniband in the case at hand).

For this approach to work, the code in the container needs to be compiled with the same version of MPI that is available in the HPC system. This is necessary because the Open MPI versions of mpiexec and orted available in the host system need to match with the compiled program. In this example the Open MPI version is v2.0.1. Therefore we need to download this version and compile it inside the container.

Note: first the example Open MPI installation that comes along with the openqcd container are removed with:

yum remove openmpi

We download Open MPI v.2.0.1 from https://www.open-mpi.org/software/ompi/v2.0 and compile it.

Openib and libibverbs need to be install to compile Open MPI over Infiniband. For that, install the epel repository on the container. This step is not required if running using TCP/IP is enough.

To install the Infiniband drivers one needs to install the epel repository.

yum install -y epel-release

The list of packages to be installed is:

openib
libibverbs, libibverbs-utils, libibverbs-devel
librdmacm, librdmacm-utils, ibacm
libnes
libibumad
libfabric, libfabric-devel
opensm-libs
swig
ibutils-libs, ibutils
opensm
libibmad
infiniband-diags

The driver needs to be installed as well, in our examples the Mellanox driver.

yum install mlx4*x86_64

The installation of both, i686 and x86_64 versions might be conflictive, and lead to an error (libibverbs: Warning: no userspace device-specific driver found for /sys/class/infiniband_verbs/uverbs0) if for example the i686 is used. The best approach is to install only the version for the architecture of the machine in this case x86_64.

The Open MPI source is compiled and installed in the container under /usr for convenience:

cd /usr
tar xvf openmpi-2.0.1.tgz
cd /usr/openmpi-2.0.1
./configure --with-verbs --prefix=/usr
make
make install

OpenQCD can then be compiled inside the udocker container in the usual way. The MPI job submission to the HPC cluster succeeds by including this line in the batch script:

/opt/cesga/openmpi/2.0.1/gcc/6.3.0/bin/mpiexec -np 128 \
  $LUSTRE/udocker-master/udocker run -e LD_LIBRARY_PATH=/usr/lib  \
  --hostenv --hostauth --user=cscdiica -v /tmp \
  --workdir=/op/projects/openQCD-1.6/main openqcd \
  /opt/projects/openQCD-1.6/main/ym1 -i ym1.in -noloc

(where $LUSTRE points to the appropriate user filesystem directory in the HPC system)

Notice that depending on the application and host operating system a variable performance degradation may occur when using the default execution mode (Pn). In this situation other execution modes (such as Fn) may provide significantly higher performance. The command udocker setup --execmode=<mode> <container-id> can be used to change between execution modes (see section 3.25).

5. Accessing GP/GPUs

The host (either the physical machine or VM) where the container will run has to have the NVIDIA driver installed. Moreover, the NVIDIA driver version has to be known apriori, since the docker image has to have the exact same version as the host

The command udocker setup --nvidia <container-id> can be used to prepare the container with the drivers necessary to run with nvidia GPGPUs. This will copy the required files from the host into the container.

Another different approach is to have docker images already prepared with the driver files but they must match what is being used in the target host. For instance base docker images with several version of the NVIDIA driver can be found in dockerhub:

In the tags tab one can check which versions are available. Dockerfiles and Ansible roles used to build these images are in the github repository: https://github.com/LIP-Computing/ansible-role-nvidia

Examples of using those NVIDIA base images with a given application are the “disvis” and “powerfit” images whose Dockerfiles and Ansible roles can be found in:

In order to build your docker image with a given CUDA or OpenCL application, the aforementioned images can be used. When the docker image with your application has been built you can run udocker with that image as described in the previous sections.

6. Accessing and transferring udocker containers

In udocker, images and containers are stored in the filesystem usually in the user home directory under $HOME/.udocker. If this location is in a shared filesystem such as in a computing farm or cluster then the content will be seen by all the hosts mounting the filesystem and can be used transparently by udocker across these hosts. If the home directory is not shared but some other location is, then you may point the UDOCKER_DIR environment variable to such a location and use it to store the udocker installation, including udocker tools, images and containers.

6.1. Directory structure

The directory structure of .udocker (or UDOCKER_DIR) is a as follows:

For a given container its directory pathname in the filesystem can be obtained as follows:

udocker inspect -p ubuntu17
/home/user01/.udocker/containers/feb0041d-e1b6-3eee-89d8-2d0617feb13a/ROOT

The pathname in the example is the root of the container filesystem tree. Below ROOT you will find all the files that comprise the container. Upon execution udocker performs a chroot like operation into this directory. You can modify, add, remove files below this location and upon execution these changes will be seen inside the container. This can be used to place or retrieve files to/from the container. By accessing this directory from the host you may also perform copies of the container directory tree e.g. for backup or other purposes.

All containers are stored under the directory “containers”. Each container is under a separate directory whose name corresponds to its alphanumeric id. This directory contains control files and the “ROOT” directory for the container filesystem.

6.2. Transfer containers with import/export or load/save

Across isolated hosts the correct way to transfer containers is to pull them from a repository such as Docker Hub. However this may implies slow downloads from remote locations and also the need to create the container again from the pulled image.

udocker provides limited support for loading images and importing containers. Containers exported to a file by Docker with docker export can be imported by udocker using:

Images saved by Docker using docker save can be imported by udocker using udocker load. Images in OCI format can also be loaded by udocker using udocker load, the format will be automatically detected.

udocker can also save images in a Docker compliant format using udocker save.

6.3. Manual transfer

The example below shows a container named MyContainer being manually transferred to another host and executed. Make sure the udocker executable is in your PATH on both the local and remote hosts.

export MYC_ROOT=$(udocker inspect -p MyContainer)
export MYC_PATH=$(dirname $MYC_ROOT)
export MYC_ID=$(basename $MYC_PATH)
export MYC_DIR=$(dirname $MYC_PATH)
cd $MYC_DIR; tar cvf - $MYC_ID | ssh user@ahost \
  "udocker install ; cd ~/.udocker/containers; tar xf -"
ssh user@ahost "udocker name $MYC_ID MyContainer; udocker run MyContainer"

7. Running as root inside containers

The behavior and capabilities of running as root inside the containers depends on the execution mode. In the Pn and Rn modes udocker will run as root. In other modes execution as root is achieved by invoking run with the --user=root option:

udocker run --user=root <container-id>

7.1. Running as root in Pn modes

In the default modes Pn, running as root is emulated, meaning that no root privileges or root capabilities are involved. The root execution is emulated by intercepting system calls and returning id 0 thus emulating a root environment.

7.2. Running as root in Fn modes

In the Fn modes running as root is not supported.

7.3. Running as root in Rn modes

The Rn (runc/crun) execution modes default to run as root, this is however achieved in a very different manner through user namespaces, as implemented by either runc or crun. These modes only work in recent Linux distributions that support user namespaces. In these execution modes the user is truly root inside the container, but with several limitations, namely on what regards access to other UIDs and GUIs. Although the user can be root inside the container, it will be a normal user outside, thus protecting the host system in case a container process breaks out. The use of user namespaces may require the setup of the system configuration files /etc/subuid and /etc/subgid which require system administrator intervention to be configured. They assign a range of UIDs and GIDs for each user to be used within the user namespaces. To overcome some of the root limitations when running inside user namespaces, udocker offers an overlay execution of proot inside runc through the execution modes R2 and R3. In these modes proot is used to overcome some of the UID and GID issues while still enabling the benefits of isolation and root execution inside de user namespaces.

7.4. Running as root in Sn modes

The Sn (singularity) execution modes default to run as normal unprivileged user. Running as “root” can be achieved with udocker run --user=root <container-id>. Execution within singularity requires namespaces and can operate in two different manners. In older distributions and kernels singularity must be installed by the system administrator with privileges. In more recent distributions and kernels singularity can operate similarly to runc and crun and take advantage of the user namespaces. In this later case UID/GID entries might also be required in /etc/subuid and /etc/subgid. Singularity is not packaged with the udocker tools tarball, but udocker can exploit existing singularity installations to run the udocker containers.

7.5. Summary of running as root

The following table provides a summary of running as root within udocker:

Mode Engine Running as root
P1 PRoot Defaults to run as root. Run as root via emulation.
P2 PRoot Same as P1
F1 Fakechroot Running as root not supported.
F2 Fakechroot Running as root not supported.
F3 Fakechroot Running as root not supported.
F4 Fakechroot Running as root not supported.
R1 runc Defaults to run as root. Run as root via user namespaces
R2 runc Same as R1 plus overlay execution with proot in mode P1.
R3 runc Same as R1 plus overlay execution with proot in mode P2.
S1 Singularity Use –user=root. Run as root via user namespaces

7.6. Running as root for software installation

Most applications and services can be run without running as root. However running as root within udocker can be useful to install software packages. Depending on the execution mode, running as root may imply additional overheads and/or security considerations.

If the software installation will need to create/change users and groups then udocker needs to run with direct access to the container passwd and group files as follows:

udocker run --user=root --containerauth <CONTAINER-ID>

For software installation the recommended execution modes are P2, S1 and R3. The emulation is not perfect and issues can still arise. Namely when using APT it can be required to install using:

apt-get -o APT::Sandbox::User=root update
apt-get -o APT::Sandbox::User=root install <package>

Upon APT errors such as cannot get security labeling handle: No such file or directory try to run as mentioned above using P2 mode, but not mounting /sys from the host by starting udocker as:

udocker.py run --user=root --nosysdirs -v /etc/resolv.conf -v /dev \
  --containerauth <CONTAINER-ID>

8. Nested execution

udocker as not been designed for nested executions, meaning execution of containers within containers. However there are successful examples of using udocker in such scenarios such as SCAR.

For running inside docker and similar: udocker offers the Fn mode which enables execution within docker or other Linux namespaces based applications.

For running udocker within udocker itself the following guidelines apply:

9. Performance

The performance experienced in the different execution modes will depend greatly on the application being executed. In general the following considerations may hold:

10. Hardware architectures

The udocker Python code has the built-in logic to support several hardware architectures namely i386, x86_64, arm (32 bit) and aarch64 (arm 64 bit). However the required engine binaries and/or libraries must also be provided for each of the architectures. Currently only some modes have compiled binaries to support execution on x86, x86_64, ARM, ARM64 and ppc64le. The executables and libraries for the execution engines shipped with udocker have a suffix that identifies the architecture, check the relevant udocker installation directories usually $HOME/.udocker/bin and $HOME/.udocker/lib.

Users may compile the same executables shipped in the udockertools in their linux hosts to support different or newer distributions, and/or architectures. See the installation manual for further information.

Checking which architectures are supported by a given container can be verified using udocker manifest inspect IMAGE. If the intended architecture is available it can be pulled using udocker pull --platform=OS/ARCH.

udocker manifest inspect centos:7
udocker pull --platform=linux/arm64 centos:7
udocker create --name=C7 centos:7
udocker run C7

In general, if the binaries in the container have been compiled for an architecture that is different from the host then the execution will not be possible. However, execution may still be possible provided that qemu-user is locally installed. In many distributions qemu-user is provided by the package qemu-user-static. In such case the default engine of udocker Pn will automatically use the qemu emulation to support the execution. Since the architecture is emulated the execution will be much slower. Emulation for the Fn modes may also work if the qemu-user binaries are both installed and also appear in /proc/sys/fs/binfmt_misc/.

11. Host environment specific notes

11.1. Termux

udocker can be used with Termux on Android, the only mode currently supported is P using PRoot. It is recommended to install and use the proot binary provided by Termux which is adapted to the Termux Android environment.

export UDOCKER_USE_PROOT_EXECUTABLE=$(which proot)
udocker run arm64v8/fedora:35

11.2. Google Colab

udocker can run on Google Colab using the P or F modes.

! pip install udocker
! udocker install
! udocker --allow-root pull centos:centos7
! udocker --allow-root create --name=c7 centos:centos7
! udocker --allow-root run c7

11.3. Docker

udocker can be used to execute containers within Docker, the only mode currently supported is F using Fakechroot.

udocker --allow-root pull ubuntu:18.04
udocker --allow-root create --name=ub18 ubuntu:18.04
udocker --allow-root setup --execmode=F3 ub18
udocker --allow-root run ub18

12. Issues

Containers should only be copied for transfer when they are in the execution modes Pn or Rn. The modes Fn perform changes to the containers that will make them fail if they are execute in a different host where the absolute pathname to the container location is different. In this later case convert back to P1 (using: udocker setup --execmode=P1) before performing the backup. Sharing of containers can be done across hosts in an homogeneous cluster or between hosts with the very same directory structure.

When experiencing issues in the default execution mode (P1) you may try to setup the container to execute using mode P2 or one of the Fn or Rn modes. See section 3.27 for information on changing execution modes.

Some execution modes require the creation of auxiliary files, directories and mount points. These can be purged from a given container using setup --purge, however this operation must be performed when the container is not being executed (nor locally nor in another host of the cluster).

Acknowledgments