Skip to content

Latest commit

 

History

History
executable file
·
662 lines (488 loc) · 25.1 KB

installation.rst

File metadata and controls

executable file
·
662 lines (488 loc) · 25.1 KB
Raw

Installing {Project}

This section will guide you through the process of installing {Project} {InstallationVersion} via several different methods. (For instructions on installing earlier versions of {Project} please see earlier versions of the docs.)

Installation on Linux

{Project} can be installed on any modern Linux distribution, on bare-metal or inside a Virtual Machine. It can even often be run nested inside another {Project} container or inside some other container system.

System Requirements

Note

This is a long section; navigation of this page is done most easily by expanding the menus on the left.

{Project} requires ~200MiB disk space once compiled and installed.

There are no specific CPU or memory requirements at runtime, though at least 2GB of RAM is recommended when building from source.

Full functionality of {Project} requires that the kernel supports:

  • FUSE - Required to mount SIF files, for fuse-overlayfs on older kernels (see OverlayFS below), and for mounting ext3 overlay filesystems. The kernels of all major Linux operating systems support FUSE by default, but system administrators must not disable it. If system administrators do not want users to mount FUSE filesystems in the primary namespace, they can remove or not install the fuse package which contains the privileged fusermount program. {Project} does not use fusermount in any mode.
  • Unprivileged user namespaces - (minimum kernel >=3.8, >=4.18 recommended) Required to run containers without root or setuid privilege. The recommended minimum version is required for unprivileged FUSE mounts.
  • OverlayFS mounts - (minimum kernel >=3.18, >=5.11 recommended) Used for creating missing bind mount paths and for writable overlays. Kernel 5.11 enables support for overlays unprivileged, but whenever the kernel OverlayFS driver doesn't work, fuse-overlayfs will be used instead.

Instructions to install without or with setuid privileges are below. Please make sure you are familiar with the discussion on Setuid & User Namespaces in the Security section of the user guide, and the :ref:`User Namespace Requirements <userns-requirements>` in this guide.

Non-standard ldconfig / Nix & Guix Environments

If {Project} is installed under a package manager such as Nix or Guix, but on top of a standard Linux distribution (e.g. RHEL or Debian), it may be unable to correctly find the libraries for --nv and --rocm GPU support. This issue occurs as the package manager supplies an alternative ldconfig, which does not identify GPU libraries installed from host packages.

To allow {Project} to locate the host (i.e. RHEL / Debian) GPU libraries correctly, set binary path in {command}.conf to point to the directory of the host's ldconfig before the $PATH: which is replaced by the user's PATH. I.E., /sbin should be before any Nix or Guix related path or the user's PATH.

Filesystem support / limitations

{Project} supports most filesystems, but there are some limitations when installing {Project} on, or running containers from, common parallel / network filesystems. In general:

  • TMPDIR / {ENVPREFIX}_TMPDIR should be on a local filesystem wherever possible.
  • If you install {Project} on a local disk on each compute node you will have the greatest chance of avoiding any limitations.
  • If {Project} is installed to a network location, it might work with the default directory layout but depending on the network filesystem type (details below) you might run into problems. The kernel overlayfs is especially selective on which filesystem types it supports, but {Project} will fall back to fuse-overlayfs if the kernel overlayfs doesn't work so you might not notice it. The issue primarily affects the directory that can be changed by the --localstatedir option to the mconfig command when compiling from source. If you have this issue, a solution may be to put most of the installation in the network location but then define --localstatedir to point to a directory that is on the local filesystem of all the nodes.

Note

Set the --localstatedir location by providing --localstatedir my/dir as an option when you configure your {Project} build with ./mconfig.

Disk usage at the --localstatedir location is negligible (<1MiB). The directory is used as a location to mount the container root filesystem, overlays, bind mounts etc. that construct the runtime view of a container. You will not see these mounts from a host shell, as they are made in a separate mount namespace. Write access to this area is not needed but a lib/{command}/mnt/session directory needs to pre-exist in it.

Overlay support

Some features of {Project}, such as the --writable-tmpfs and --overlay options as well as creation of missing bind mountpoints, use overlay mounts to construct a container root filesystem that combines files from different locations. Overlay mounts may use the Linux kernel overlay filesystem driver or the fuse-overlayfs userspace implementation, depending on the workflow and support from the host kernel.

Overlay support has two aspects, referenced below:

  • lowerdir support for a filesystem allows a directory on that filesystem to act as the 'base' of a container. A filesystem must support overlay lowerdir for you be able to run {aProject} sandbox container on it, while using functionality such as --writable-tmpfs / --overlay.
  • upperdir support for a filesystem allows a directory on that filesystem to be merged on top of a lowerdir to construct a container. If you use the --overlay option to overlay a directory onto a container, then the filesystem holding the overlay directory must support upperdir.

Overlays are mounted with the Linux kernel driver when:

  • the upperdir is not an extfs image, and
  • the upperdir is not a directory on an unsupported filesystem type (detailed later on this page), and
  • {Project} is running either in setuid mode or in unprivileged / non-setuid mode when the kernel supports unprivileged overlay mounts.

Overlays are mounted with the fuse-overlayfs userspace implementation in all other cases.

Fakeroot with uid/gid mapping on Network filesystems

When {Project} is run using the :ref:`fakeroot <fakeroot>` option and mappings are available in /etc/subuid and /etc/subgid then uids / gids inside the container are mapped to different host uids / gids.

Most local filesystems (ext4/xfs etc.) have no problem with this uid/gid mapping. Most network filesystems (NFS/Lustre/GPFS etc.), however, only support a single uid for each user. When additional user ids are attempted to be used the fileserver will deny many operations, with 'permission denied' errors. This is currently a generic problem for rootless container runtimes.

This is only a problem when sandbox images are stored on the network filesystems. For that case it is probably better to use one of the other fakeroot modes that {Project} supports. Alternatively, use SIF images instead of sandbox images because they don't have the problem.

{Project} cache / atomic rename

{Project} will cache SIF container images generated from remote sources, and any OCI/docker layers used to create them. The cache is created at $HOME/.{command}/cache by default. The location of the cache can be changed by setting the {ENVPREFIX}_CACHEDIR environment variable.

The directory used for {ENVPREFIX}_CACHEDIR should be:

  • A unique location for each user. Permissions are set on the cache so that private images cached for one user are not exposed to another. This means that {ENVPREFIX}_CACHEDIR cannot be shared.
  • Located on a filesystem with sufficient space for the number and size of container images anticipated.
  • Located on a filesystem that supports atomic rename, if possible.

The {Project} cache is concurrency safe. Parallel runs of {Project} that would create overlapping cache entries will not conflict, as long as the filesystem used by {ENVPREFIX}_CACHEDIR supports atomic rename operations.

Support for atomic rename operations is expected on local POSIX filesystems, but varies for network / parallel filesystems and may be affected by topology and configuration. For example, Lustre supports atomic rename of files only on a single MDT. Rename on NFS is only atomic to a single client, not across systems accessing the same NFS share.

If you are not certain that your $HOME or {ENVPREFIX}_CACHEDIR filesystems support atomic rename, do not run {command} in parallel using remote container URLs. Instead use {command} pull to create a local SIF image, and then run this SIF image in a parallel step. An alternative is to use the --disable-cache option, but this will result in each {Project} instance independently fetching the container from the remote source, into a temporary location.

NFS

NFS filesystems support overlay mounts as a lowerdir only, and do not support --fakeroot.

  • Containers run from SIF files located on an NFS filesystem do not have restrictions.
  • You cannot use --overlay to overlay a writable directory onto a container when the overlay (upperdir) directory is on an NFS filesystem. It only works with a read-only overlay.
  • When building a container, or running a container with --fakeroot, your TMPDIR / {ENVPREFIX}_TMPDIR should not be set to an NFS location.
  • You should not run a sandbox container with --fakeroot from an NFS location.
Lustre / GPFS / PanFS

Lustre, GPFS, and PanFS do not have sufficient upperdir or lowerdir overlay support for certain {Project} features, and do not support --fakeroot.

  • --overlay or --writable-tmpfs with a sandbox container (lowerdir) that is located on a Lustre, GPFS, or PanFS filesystem will use fuse-overlayfs. SIF containers on Lustre, GPFS, and PanFS will work correctly with the same overlay method as on any filesystem type.
  • --overlay with an overlay (upperdir) directory on a Lustre, GPFS, or PanFS filesystem will use fuse-overlayfs, not the kernel overlay system.
  • When building a container, or running a container with --fakeroot, your TMPDIR / {ENVPREFIX}_TMPDIR should not be a Lustre, GPFS, or PanFS location.
FUSE-based filesystems

The kernel overlay driver does not allow the upperdir to be a FUSE-based filesystem, so those always use fuse-overlayfs. A FUSE-based filesystem is used by default when the upperdir is an extfs image.

Install unprivileged from pre-built binaries

A script is available to make a relocatable unprivileged binary installation of {Project}, including all dependencies that are not normally installed on Linux systems. The script works on current Red Hat Enterprise Linux-derived systems, Fedora, SUSE/OpenSUSE, Debian, and Ubuntu.

This is the easiest way to use it:

$ curl -s https://raw.githubusercontent.com/{orgrepo}/main/tools/install-unprivileged.sh | \
    bash -s - install-dir

where install-dir is the directory you want to install it into. Once installed, you can run it with install-dir/bin/{command}.

The script requires a few tools to be in the user's PATH that aren't always available: curl, rpm2cpio, and cpio. If they are not available and you do not have privileges to install them, you can copy them from another compatible machine or you can run the script on another machine and copy the installed files to the machine you want them on.

Multiple architectures can be installed into the same directory on a shared filesystem and it will automatically pick the correct architecture when it runs. Architectures supported by EPEL and Fedora are supported. By default it installs from the oldest supported EPEL version available, for maximum compatibility when working with containers based on old or new operating system distributions.

Install from pre-built packages

Prebuilt packages are available for released versions of {Project} on a variety of host operating systems.

Install RPM from EPEL or Fedora

Multiple architectures of RPMs are available for Red Hat Enterprise Linux and Fedora.

First, on Red Hat Enterprise Linux derived systems enable the EPEL repositories like this:

$ sudo dnf install -y epel-release

Then to install a non-setuid installation of {Project} do:

$ sudo dnf install -y {command}

or for a setuid installation do:

$ sudo dnf install -y {command}-suid

Install from GitHub release RPMs

Alternatively, x86_64 RPMs are available on GitHub immediately after each {Project} release and they can be installed directly from there. For the non-setuid installation:

$ sudo dnf install -y https://github.com/{orgrepo}/releases/download/v{InstallationVersion}/{command}-{GitHubDownloadVersion}.x86_64.rpm

For the setuid installation do above command first and then this one:

$ sudo dnf install -y https://github.com/{orgrepo}/releases/download/v{InstallationVersion}/{command}-suid-{GitHubDownloadVersion}.x86_64.rpm

Install Debian packages

Pre-built Debian packages are only available on GitHub and only for the amd64 architecture. For the non-setuid installation use these commands:

$ sudo apt update
$ sudo apt install -y wget
$ cd /tmp
$ wget https://github.com/{orgrepo}/releases/download/v{InstallationVersion}/{command}_{InstallationVersion}_amd64.deb
$ sudo apt install -y ./{command}_{InstallationVersion}_amd64.deb

For the setuid installation do above commands first and then these:

$ wget https://github.com/{orgrepo}/releases/download/v{InstallationVersion}/{command}-suid_{InstallationVersion}_amd64.deb
$ sudo dpkg -i ./{command}-suid_{InstallationVersion}_amd64.deb

Install Ubuntu packages

Pre-built Ubuntu packages are available on PPA (Personal Package Archive) for the amd64 and arm64 architectures on all current Ubuntu releases.

First, on Ubuntu based containers install software-properties-common package to obtain add-apt-repository command. On Ubuntu Desktop/Server derived systems skip this step.

$ sudo apt update
$ sudo apt install -y software-properties-common

For the non-setuid installation use these commands:

$ sudo add-apt-repository -y ppa:apptainer/ppa
$ sudo apt update
$ sudo apt install -y apptainer

For the setuid installation do above commands first and then these:

$ sudo add-apt-repository -y ppa:apptainer/ppa
$ sudo apt update
$ sudo apt install -y apptainer-suid

Install from Source

To install from source, follow the instructions in INSTALL.md on GitHub.

Relocatable Installation

An unprivileged (non-setuid) {Project} installation built from source is relocatable. As long as the structure inside the installation directory (--prefix) is maintained, it can be moved to a different location and {Project} will continue to run normally.

Relocation of a setuid installation is not supported, as restricted location / ownership of configuration files is important to security.

Source bash completion file

If you installed from source, then to enjoy bash shell completion with {Project} commands and options, source the bash completion file (assuming the default installation prefix):

$ . /usr/local/share/bash-completion/completions/{command}

Add this command to your ~/.bashrc file so that bash completion continues to work in new shells. (Adjust the path if you installed {Project} to a different location.)

Build an RPM

If you use RHEL, a RHEL derivate, or SUSE, building and installing {aProject} RPM allows your {Project} installation be more easily managed, upgraded and removed.

The instructions on how to build the RPM from source are in a INSTALL.md section on GitHub.

Build a Debian package

Packaging for Debian and Ubuntu can also be built from source. The instructions on how to do that are in a separate file DEBIAN_PACKAGE.md on GitHub.

Testing & Checking the Build Configuration

After installation you can perform a basic test of {Project} functionality by executing a simple alpine container:

$ {command} exec docker://alpine cat /etc/alpine-release
3.9.2

See the user guide for more information about how to use {Project}.

{command} buildcfg

Running {command} buildcfg will show the build configuration of an installed version of {Project}, and lists the paths used by {Project}. Use {command} buildcfg to confirm paths are set correctly for your installation, and troubleshoot any 'not-found' errors at runtime.

$ {command} buildcfg
PACKAGE_NAME={command}
PACKAGE_VERSION={InstallationVersion}
BUILDDIR=/home/dtrudg/Sylabs/Git/{command}/builddir
PREFIX=/usr/local
EXECPREFIX=/usr/local
BINDIR=/usr/local/bin
SBINDIR=/usr/local/sbin
LIBEXECDIR=/usr/local/libexec
DATAROOTDIR=/usr/local/share
DATADIR=/usr/local/share
SYSCONFDIR=/usr/local/etc
SHAREDSTATEDIR=/usr/local/com
LOCALSTATEDIR=/usr/local/var
RUNSTATEDIR=/usr/local/var/run
INCLUDEDIR=/usr/local/include
DOCDIR=/usr/local/share/doc/{command}
INFODIR=/usr/local/share/info
LIBDIR=/usr/local/lib
LOCALEDIR=/usr/local/share/locale
MANDIR=/usr/local/share/man
{ENVPREFIX}_CONFDIR=/usr/local/etc/{command}
SESSIONDIR=/usr/local/var/{command}/mnt/session
PLUGIN_ROOTDIR=/usr/local/libexec/apptainer/plugin
APPTAINER_CONF_FILE=/usr/local/etc/apptainer/apptainer.conf
APPTAINER_SUID_INSTALL=0

Note that the LOCALSTATEDIR and SESSIONDIR should be on local, non-shared storage.

Test Suite

The {Project} codebase includes a test suite that is run during development using CI services.

If you would like to run the test suite locally you can run the test targets from the builddir directory in the source tree:

  • make check runs source code linting and dependency checks
  • make unit-test runs basic unit tests
  • make integration-test runs integration tests
  • make e2e-test runs end-to-end tests, which exercise a large number of operations by calling the {Project} CLI with different execution profiles.

Note

Running the full test suite requires a docker installation, and nc in order to test docker and instance/networking functionality.

{Project} must be installed with suid in order to run the full test suite, as it must run the CLI with setuid privilege for the starter-suid binary.

Warning

sudo privilege is required to run the full tests, and you should not run the tests on a production system. We recommend running the tests in an isolated development or build environment.

Installation on Windows or Mac

Linux container runtimes like {Project} cannot run natively on Windows or Mac because of basic incompatibilities with the host kernel. (Contrary to a popular misconception, macOS does not run on a Linux kernel. It runs on a kernel called Darwin originally forked from BSD.)

To run {Project} on a Windows or macOS computer, a Linux virtual machine (VM) is required. There are various ways to configure a VM on both Windows and macOS. On Windows, we recommend the Windows Subsystem for Linux (WSL2), and on macOS, we recommend Lima.

Windows

Recent builds of Windows 10, and all builds of Windows 11, include version 2 of the Windows Subsystem for Linux. WSL2 provides a Linux virtual machine that is tightly integrated with the Windows environment. The default Linux distribution used by WSL2 is Ubuntu. It is straightforward to install {Project} inside WSL2 Ubuntu, and use all of its features.

Follow the WSL2 installation instructions to enable WSL2 with the default Ubuntu 22.04 environment. On Windows 11 and the most recent builds of Windows 10 this is as easy as opening an administrator command prompt or Powershell window and entering:

wsl --install

Follow the prompts. A restart is required, and when you open the 'Ubuntu' app for the first time you'll be asked to set a username and password for the Linux environment.

You can then install {Project} from source, from the Debian packages on the {Project} release page, or from the Ubuntu PPA.

GPU Support

WSL2 supports using an NVIDIA GPU from the Linux environment. To use a GPU from {Project} in WSL2, you must first install libnvidia-container-tools, following the instructions in the libnvidia-container documentation:

curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg \
curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | \
  sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \
  sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list \
sudo apt-get update
sudo apt-get install -y nvidia-container-toolkit

Once this process has been completed, GPU containers can be run under WSL2 using the --nv and --nvccli flags together:

$ {command} pull docker://tensorflow/tensorflow:latest-gpu

$  {command} run --nv --nvccli tensorflow_latest-gpu.sif
INFO:    Setting 'NVIDIA_VISIBLE_DEVICES=all' to emulate legacy GPU binding.
INFO:    Setting --writable-tmpfs (required by nvidia-container-cli)
________                               _______________
___  __/__________________________________  ____/__  /________      __
__  /  _  _ \_  __ \_  ___/  __ \_  ___/_  /_   __  /_  __ \_ | /| / /
_  /   /  __/  / / /(__  )/ /_/ /  /   _  __/   _  / / /_/ /_ |/ |/ /
/_/    \___//_/ /_//____/ \____//_/    /_/      /_/  \____/____/|__/
You are running this container as user with ID 1000 and group 1000,
which should map to the ID and group for your user on the Docker host. Great!
{Project}> python
Python 3.8.10 (default, Nov 26 2021, 20:14:08)
[GCC 9.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> tf.config.list_physical_devices('GPU')
2022-03-25 11:42:25.672088: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:01:00.0/numa_node
Your kernel may have been built without NUMA support.
2022-03-25 11:42:25.713295: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:01:00.0/numa_node
Your kernel may have been built without NUMA support.
2022-03-25 11:42:25.713892: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:922] could not open file to read NUMA node: /sys/bus/pci/devices/0000:01:00.0/numa_node
Your kernel may have been built without NUMA support.
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

Note that the --nvccli flag is required to enable container setup using the nvidia-container-cli utility. {Project}'s simpler library binding approach (--nv only) is not sufficient for GPU support under WSL2.

Mac

{Project} is available via Lima (installable with Homebrew or manually)

To use Lima via Homebrew:

$ /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"
$ brew install qemu lima

Then do limactl start template://apptainer and limactl shell apptainer.

To customize cpus and memory on the VM, you can use --cpus 1 and --memory 1.

See the lima "apptainer" template and the lima "default" template for more details.

By default, the host home directory is mounted as read-only in the guest, but there is also a shared writable directory mounted in /tmp/lima that can be accessed both from the host and in the guest.