SingularityCE 3.9.0 Release Candidate 1

This is the first release candidate for the upcoming SingularityCE 3.9.0. We'd be grateful all testing, bug reports, and comments, as we look forward to a stable 3.9.0 release.

Various behavior changes and new features have been introduced. Please carefully review the release notes below, and refer to the 'master branch (unreleased)' documentation at https://sylabs.io/docs/

Changed defaults / behaviours

• Building SingularityCE 3.9.0 requires go >=1.16. We now aim to support the two most recent stable versions of Go. This corresponds to the Go Release Maintenance Policy and Security Policy, ensuring critical bug fixes and security patches are available for all supported language versions.
• LABELs from Docker/OCI images are now inherited. This fixes a longstanding regression from Singularity 2.x. Note that you will now need to use --force in a build to override a label that already exists in the source Docker/OCI container.
• The source paths for %files lines in a definition file are no longer interpreted by a shell. This means that environment variable substitution is not performed. Previously, environment variables were substituted for source paths, but not destination paths, leading to unexpected copy behaviour. Globbing for source files will now follow the Go filepath.Match pattern
  syntax.
• Removed --nonet flag, which was intended to disable networking for in-VM execution, but has no effect.
• --nohttps flag has been deprecated in favour of --no-https. The old flag is still accepted, but will display a deprecation warning.
• Paths for cryptsetup, go, ldconfig, mksquashfs, nvidia-container-cli, unsquashfs are now found at build time by mconfig and written into singularity.conf. The path to these executables can be overridden by changing the value in singularity.conf. If the path is not set in singularity.conf then the the executable will be found by searching $PATH.
• When calling ldconfig to find GPU libraries, singularity will not fall back to /sbin/ldconfig if the ldconfig on $PATH errors. If installing in a Guix/Nix on environment on top of a standard host distribution you must set ldconfig path = /sbin/ldconfig to use the host distribution ldconfig to find GPU libraries.
• --nv will not call nvidia-container-cli to find host libraries, unless the new experimental GPU setup flow that employs nvidia-container-cli for all GPU related operations is enabled (see below).
• If a container is run with --nvcli and --contain, only GPU devices specified via the NVIDIA_VISIBLE_DEVICES environment variable will be exposed within the container. Use NVIDIA_VISIBLE_DEVICES=all to access all GPUs inside a container run with --nvccli.
• Example log-plugin rewritten as a CLI callback that can log all commands executed, instead of only container execution, and has access to command arguments.
• An invalid remote build source (bootstrap) will be identified before attempting to submit the build.
• The bundled reference CNI plugins are updated to v1.0.1. The flannel plugin is no longer included, as it is maintained as a separate plugin at: https://github.com/flannel-io/cni-plugin. If you use the flannel CNI plugin you should install it from this repository.
• Instances are no longer created with an IPC namespace by default. An IPC namespace can be specified with the -i|--ipc flag.

New features / functionalities

• --writable-tmpfs can be used with singularity build to run the %test section of the build with a ephemeral tmpfs overlay, permitting tests that write to the container filesystem.
• --compat flag for actions is a new short-hand to enable a number of options that increase OCI/Docker compatibility. Infers --containall, --no-init, --no-umask, --writable-tmpfs. Does not use user, uts, or network namespaces as these may not be supported on many installations.
• --no-https now applies to connections made to library services specified in --library://<hostname>/... URIs.
• remote add --insecure may be used to configure endpoints that are only accessible via http.
• The experimental --nvccli flag will use nvidia-container-cli to setup the container for Nvidia GPU operation. SingularityCE will not bind GPU libraries itself. Environment variables that are used with Nvidia's docker-nvidia runtime to configure GPU visibility / driver capabilities & requirements are parsed by the --nvccli flag from the environment of the calling user. By
  default, the compute and utility GPU capabilities are configured. The use nvidia-container-cli option in singularity.conf can be set to yes to always use nvidia-container-cli when supported. Note that in a setuid
  install, nvidia-container-cli will be run as root with required ambient capabilities. --nvccli is not currently supported in the hybrid fakeroot (setuid install + --fakeroot) workflow. Please see documentation for more details.
• The --apply-cgroups flag can be used to apply cgroups resource and device restrictions on a system using the v2 unified cgroups hierarchy. The resource restrictions must still be specified in the v1 / OCI format, which will be translated into v2 cgroups resource restrictions, and eBPF device restrictions.
• A new --mount flag and SINGULARITY_MOUNT environment variable can be used to specify bind mounts in type=bind,source=<src>,destination=<dst>[,options...] format. This improves CLI compatibility with other runtimes, and allows binding paths containing : and , characters (using CSV style escaping).

Bug fixes

• The oci commands will operate on systems that use the v2 unified cgroups hierarchy.

Thanks / Reporting Bugs

Thanks to our contributors for code, feedback and, testing efforts!

As always, please report any bugs to: https://github.com/sylabs/singularity/issues/new

If you think that you've discovered a security vulnerability please report it to: [email protected]

Have fun!

Downloads

Please use the singularity-ce-3.9.0-rc.1.tar.gz download below to obtain and install SingularityCE 3.9.0-rc.1. The GitHub auto-generated 'Source Code' downloads do not include required dependencies etc.