SPACK Mirror on JLab CUE

From epsciwiki
Jump to navigation Jump to search

Using the JLab SPACK Repository

Overview

SPACK is a package manager used to maintain multiple versions of software compiled with various compilers for various OSes. The EPSCI group takes the primary responsibility for maintaining the SPACK repository at JLab. SPACK has a rich feature set that allows a lot of flexibility in how one can use it to manage their software. This page describes details of how SPACK is implemented at JLab for the ENP program.

There are three primary use cases for the software built with the SPACK system:

  1. Users on the JLab CUE want to use the pre-built binary versions on JLab computers
  2. Users running offsite want to use the pre-built binary versions on their local computers
  3. Users want to install the pre-built binaries on their local computer so they can run untethered

The first two of these are satisfied by using /cvmfs. The third use case uses a web accessible SPACK buildcache and is quite a bit more fickle. Officially, we do not support option 3 because of this.

Quickstart

The recommended way to set up your environment is with one of the following:

  [bash]  source /cvmfs/oasis.opensciencegrid.org/jlab/epsci/spack_env/spack_env.sh  lmod gcc/9.3.0
  [tcsh]  source /cvmfs/oasis.opensciencegrid.org/jlab/epsci/spack_env/spack_env.csh lmod gcc/9.3.0

Note that the above may take a few seconds to complete, but it sets up a user-friendly package naming scheme for "module load". If you want quicker startup and are willing to live with package names that include long hashes, then source the script with no arguments:

  [bash]  source /cvmfs/oasis.opensciencegrid.org/jlab/epsci/spack_env/spack_env.sh
  [tcsh]  source /cvmfs/oasis.opensciencegrid.org/jlab/epsci/spack_env/spack_env.csh

Other useful commands:

  module avail                # List available packages
  module load packagename     # Load a package (optionally specify version number)
  module unload packagename   # Unload a package that was previously loaded

The following operating systems are supported:

Supported Operating Systems
OS support start date support end date
centos/7.7.1908 March 31, 2021 current
centos/8.3.2011 March 31, 2021 current
ubuntu/21.04 March 31, 2021 current

CVMFS Client Configuration

If you are working on the JLab ifarm computers than CVMFS is already installed and configured. This is nothing else you need to do. CVMFS may also already be available on many remote HPC sites (e.g. NERSC). Check the site's specific documentation or simply look for the /cvmfs/oasis.opensciencegrid.org directory.

To mount the public, read-only CVMFS volume that contains the pre-built binaries see the instructions in one of the following sections for your specific platform.

The most up to date instructions on installing and configuring the CVMFS client software can be found on the CVMFS website.


Linux

Here are instructions for installing on a CentOS or RedHat system (personal laptop or desktop)

1. Install the pointer to the CVMFS repo and then install cvmfs itself. After it is installed, generate a default config file.

  sudo yum install https://ecsft.cern.ch/dist/cvmfs/cvmfs-release/cvmfs-release-latest.noarch.rpm
  sudo yum install -y cvmfs
  cvmfs_config setup

2. Create a config file /etc/cvmfs/default.local with the following content (you need to do this with sudo):

 CVMFS_REPOSITORIES=oasis.opensciencegrid.org
 CVMFS_HTTP_PROXY=DIRECT
 CVMFS_CLIENT_PROFILE=single

3. Restart the autofs service

  systemctl restart autofs

Mac OS X

To use CVMFS on Mac OS X, you need to install the MacFUSE package and then the cvmfs package. You should then reboot so everything will load properly. The step-by-step instructions follow.

1. Download and install the macFUSE package

2. Download and install the cvmfs package with the following (Note that downloading the cvmfs package via curl apparently avoids some signature security issue on Mac OS X that you would get if downloaded via web-browser. Don't ask me how.)

 curl -o ~/Downloads/cvmfs-2.7.5.pkg https://ecsft.cern.ch/dist/cvmfs/cvmfs-2.7.5/cvmfs-2.7.5.pkg
 open cvmfs-2.7.5.pkg

3. Create a config file /etc/cvmfs/default.local with the following content (you need to do this with sudo):

 CVMFS_REPOSITORIES=oasis.opensciencegrid.org
 CVMFS_HTTP_PROXY=DIRECT

4. Restart the computer

5. Create the mount point and mount oasis with:

 sudo mkdir -p /cvmfs/oasis.opensciencegrid.org
 sudo mount -t cvmfs oasis.opensciencegrid.org /cvmfs/oasis.opensciencegrid.org

If it all works you should see something like this:

 >sudo mount -t cvmfs oasis.opensciencegrid.org /cvmfs/oasis.opensciencegrid.org
 CernVM-FS: running with credentials 10000:10000
 CernVM-FS: loading Fuse module... done
 CernVM-FS: mounted cvmfs on /Users/Shared/cvmfs/oasis.opensciencegrid.org

Docker

There are actually two options for using CVMFS inside a Docker container:

  1. Install CVMFS on the host and simply bind the /cvmfs directory to the same directory inside the container
  2. Run the CVMFS software inside the container and mount it there.

Option 1 is preferred since any caching of the files is done by the host and and so does not disappear when the container goes away. It also can be used with any image and does not require another image to be created with the CVMFS software installed. To implement option 1, first mount CMVFS on the host using the above instructions for your host platform. Then, when you start the container, give the docker command an argument of -v /cvmfs:/cvmfs.

Option 2 can be convenient if you have trouble getting CVMFS working on the host. There are actually two methods here. One is to use the pre-made Docker container as described in the CVMFS documentation. You may create an image based on this or even use it as-is to supply /cvmfs to the host and then use option 1 above.

The second method is to create a new image from scratch containing the necessary software. This method has worked in the past, though there may be easier ways of doing it today. Here are the instructions though in case all of the other above methods fail.

Unfortunately, there are a couple of steps that cannot be done when the image is created and must be implemented when the container is created. A working example with some comments can be found here:

https://github.com/faustus123/hdcontainers/tree/master/Docker_cvmfs


Running untethered (no CVMFS)

Running untethered means installing the packages on you local computer so you can still run the software even with no internet connection. It is stated up front that this is unlikely to work for numerous reasons, but for those who like punishing themselves, here is some info that may help get you going. It goes without saying that none of this is recommended.

The main issue with installing locally is that many packages build their installation paths into their installed scripts and binaries. While spack does have mechanisms to try and fix this, they can fail if the directory path is either too long or too short. Your best chances of success will come if you create a local directory path that matches exactly what it would be if /CVMFS were mounted. Here are some example instructions:

  mkdir -p /cvmfs/oasis.opensciencegrid.org/jlab/epsci/centos/
  git clone --depth 1 https://github.com/spack/spack.git /cvmfs/oasis.opensciencegrid.org/jlab/epsci/centos/7.7.1908
  source /cvmfs/oasis.opensciencegrid.org/jlab/epsci/centos/7.7.1908/share/spack/setup-env.sh # or setup-env.csh
  spack mirror add jlab-public https://spack.jlab.org/mirror
  spack install -f -o -u clhep   # This should install "CLHEP" locally using the pre-built binaries 




Administration of the SPACK Repository

The following sections describe various aspects of creating and managing the JLab SPACK repository. There are a number of choices that were made in how this was set up so this documents those since they may not all be obvious by simply looking at directory structures and config. files.

Perhaps one of the most important pieces of information is that scripts and tools used to help us maintain spack at JLab are kept in a github repository:

  https://github.com/JeffersonLab/epsci-spack/tree/main/admin

which is checked out on the CUE in /scigroup/spack/admin.

Organizational Overview

The organization of the spack binaries is as follows:

  1. Packages are built using singularity containers
    • Containers bind the /scigroup/cvmfs subdirectory to be at /cvmfs/oasis.opensciencegrid.org/jlab inside the container
    • This allows absolute paths that start with /cvmfs to be writable in the build/install process
    • The /scigroup/cvmfs/epsci directory is exported to CVMFS so it can be mounted read-only from anywhere
    • The export is done every 4 hours via cronjob. Thus, newly built packages will not be immediately accessible.
      • Wes Moore set this up and can increase frequency if needed.
  2. A separate spack repository is maintained for every platform (e.g. centos/7.7.1908 is separate from centos/8.0.2011)
    • This was a choice made on our end to segregate the binaries and make it easier to add and drop support for platforms in the future.
  3. In addition to the global spack repository, we also include the eic-spack and epsci-spack repositories.
    • This allows us to pull from the eic-spack package configurations and maintain our own package configurations.
  4. Users will access the software via the /cvmfs directory.
    • The SciComp computers (e.g. ifarm1901) all mount /cvmfs
    • Users can also install the CVMFS client on their personal laptop or desktop to access the software.
  5. The packages are exported to a build cache accessible from https://spack.jlab.org/mirror
    • We do this only because it is simple and doesn't cost us anything significant. We discourage its use and may remove it in the future.

Setting up a new platform

A platform here is defined as the OS name and version. e.g. centos/7.7.1908. The specific versions are chosen based on official Docker images maintained by the OS vendors on Docker Hub(for examples, look here). For platforms corresponding to CUE machines, the exact tags used are selected to be as close as possible to what is being used on the CUE.

The basic steps are to create a singularity image, then use it to setup a new spack instance ( helper script is available for this). Packages will then be built with the native compiler for the platform. Optionally, other compiler versions may be built using spack and then those compilers used to build versions of the spack packages compatible with that compiler.

The following sections describe these steps in some detail.

Creating a new Singularity Image

For the purposes of this system, the Singularity images used for building packages are derived from Docker images. This ensures that either Docker or Singularity can be used to build packages with spack. Thus, if someone needs a convenient sandbox to work with locally they can choose the container system that is most convenient for them. Docker images we create are posted on Docker Hub where Singularity can easily pull them. (Docker images cannot be easily created from Singularity images.)

The Dockerfiles used to create the Docker images are kept in the git-hub repository "epsci-containers". They are also copied into the image itself so one can always access the Dockerfile used to create an image via /container/Dockerfile.*. The Docker images are created with only a few system software packages installed. Mainly a C++ compiler, version control tools (e.g. git and svn), python, and a couple of other tools needed for building packages. Below is an example of a Dockerfile (click right-hand side to view).

EXAMPLE Dockerfile. (Click "Expand" to the right to see the example -->):

#--------------------------------------------------------------------------
# ubuntu build environment
# 
# This Dockerfile will produce an image based on the one used for running
# at NERSC, PSC, and the OSG, but which can also be used to mount CVMFS
# using any computer. The main use case is to provide a simple way to
# mount and run software from /group/halld/Software on you local laptop
# or desktop.
#
# To use this most effectively:
#
#      docker run -it --rm jeffersonlab/epsci-ubuntu cat /container/dsh | tr -d "\r" > dsh
#      chmod +x ./dsh
#      ./dsh jeffersonlab/epsci-ubuntu
#
#--------------------------------------------------------------------------
#
#   docker build -t epsci-ubuntu:21.04 -t jeffersonlab/epsci-ubuntu:21.04 .
#   docker push jeffersonlab/epsci-ubuntu:21.04
#
#--------------------------------------------------------------------------   

FROM ubuntu:21.04
 
# Python3 requires the timezone be set and will try and prompt for it.
ENV TZ=US/Eastern
RUN ln -snf /usr/share/zoneinfo/$TZ /etc/localtime && echo $TZ > /etc/timezone

# Install compiler and code management tools
RUN apt -y update \
	&& apt -y install build-essential libssl-dev libffi-dev python-dev \
	&& apt -y install python python3 git subversion cvs curl

COPY dsh /container/dsh
COPY Dockerfile /container/Dockerfile
RUN ln -s /root /home/root
RUN ln -s /root /home/0

CMD ["/bin/bash"]


To create a singularity image, one first needs to create a Docker image. Thus, one needs access to a computer with Docker installed. This generally needs to be a personal desktop or laptop since Docker requires root access and is therefore not available on the public machines like ifarm. (Incidentally, singularity also requires root privileges in order to build an image from a recipe, but not if just pulling from an existing Docker image). Here is example of the steps you might go through if creating an image for a different version of ubuntu. This assumes you are starting on a computer with Docker installed and running.

  1. git clone https://github.com/JeffersonLab/epsci-containers
  2. cd epsci-containers/base
  3. cp Dockerfile.ubuntu.21.04 Dockerfile.ubuntu.18.04
  4. edit Dockerfile.ubuntu.18.04 to replace the version numbers with the new ones. They appear in a lot of places so better to do global replace
  5. docker build -t epsci-ubuntu:18.04 -t jeffersonlab/epsci-ubuntu:18.04 -f Dockerfile.ubuntu.18.04 .
  6. docker push jeffersonlab/epsci-ubuntu:18.04
  7. ssh ifarm
  8. module use /apps/modulefiles
  9. module load singularity
  10. cd /scigroup/spack/mirror/singularity/images
  11. singularity build epsci-ubuntu-18.04.img docker://jeffersonlab/epsci-ubuntu:18.04

Initializing the new platform configuration

There are some pitfalls that are easy to fall into when trying to setup a new platform. Particularly if you want to build using a non-default compiler. To ameliorate this several administration scripts have been written to make it as turnkey as possible.

Step-by-step instructions are below, but you may find some useful details in the comments at the top of the script: /scigroup/spack/admin/mnp.sh

To set up the initial directory and build some of the base packages do the following. Note that this assumes a singularity container exists in the standard location for the platform version you are setting up (see [#Singularity|singularity section above] for details).

In this example we assume we are building for a new platform named "rocky/8".

  1. cd /scigroup/spack/admin
  2. cp make_new_platform_centos7.sh make_new_platform_rocky8.sh
  3. <edit the settings at the top of the new make_new_platform_rocky8.sh script>
  4. ./make_new_platform_rocky8.sh

To add a list of packages to the spack instance, write them to a text file with one line per package specification. An example can be seen here: n.b. this can also be done via a yaml file which may eventually be used to replace this system

  jlabce-2.4.txt
  1. cd /scigroup/spack/admin
  2. cp add_to_platform_centos7.sh add_to_platform_rocky8.sh
  3. <edit the settings at the top of the new add_to_platform_rocky8.sh script>
  4. ./add_to_platform_rocky8.sh jlabce-2.4.txt


Setting up the module system (LMOD)

The mnp.sh script should already set up the LMOD system configuration when a new platform is created. This section documents some of what was done and why.

We would like most users to be be able to interact with the spack packages using the standard "module load" command. Spack has nice support for this though there are options for how it is setup and we'd like to be consistent across supported platforms.

First off, we use the LMOD system as it supports hierarchical module files. This allows us to configure the system so that when a specific compiler is loaded, only packages corresponding to that compiler are listed. This should make it easier on the user to navigate and to avoid loading incompatible packages. We also configure it to present packages using the {package}/{version} naming scheme. This is what is used by /apps on the CUE which will make the spack packages integrate more seamlessly with those.

Modules configuration file

The ${SPACK_ROOT}/etc/spack/modules.yaml configuration file must be created and have the following content added. This is mostly based on an example given in the spack documentation under "Hierarchical Module Files". Descriptions of the settings are given below.

EXAMPLE modules.yaml. (Click "Expand" to the right to see the example -->):

 modules:
   enable::
     - lmod
   lmod:
     core_compilers:
       - 'gcc@4.8.5'
     hierarchy:
       - mpi
     hash_length: 0
     whitelist:
       - gcc
     blacklist:
       - '%gcc@4.8.5'
       - 'arch=linux-centos7-zen2'
     all:
       filter:
         environment_blacklist:
           - "C_INCLUDE_PATH"
           - "CPLUS_INCLUDE_PATH"
           - "LIBRARY_PATH"
       environment:
         set:
           '{name}_ROOT': '{prefix}'
     projections:
       all:          '{name}/{version}'
       ^lapack:      '{name}/{version}-{^lapack.name}'
  • The core_compilers section should list the system compiler as the default.
  • hash_length: 0 removes the spack hash from package names
  • whitelist ensures all gcc compilers are available. (Once one of those is loaded, other packages will appear.)
  • blacklist excludes packages built with the default system compiler (n.b. whitelist overrides this so other compilers will still be listed)
    • the arch= blacklist line excludes packages built specifically with the zen2 microcode instead of generic x86_64. Those packages were actually built by mistake and may be removed altogether. This is a nice way though of obscuring them from view.
  • The environment_blacklist filter was just copied from the spack example in the documentation. We may want to remove it. I did not recall build systems using those variables so just left it in.
  • The environment: set: section adds an environment variable for every package that is {package}_ROOT so the root directory of the package can be easily obtained, even if the package itself does not define such a variable.
  • The projections section defines the module naming scheme. The line for lapack was left in from the spack tutorial example.

Building a new package

We anticipate getting user requests for new packages. Building packages that already have a spack configuration should be fairly straight-forward. There are a couple of questions to answer before you do it though are:

  1. Is this package something we should be supporting via spack?
  2. Should this be part of an Environment?
  3. What compilers/platforms should this be built for?

The best way to handle this is to use the appropriate add_to_platform_X.sh script with the exact package specification written in a text file. If it is not being added to an existing spack Environment then it should be added to the misc_packages.txt file for archival purposes. Make sure to commit any changes to github.

Building a new package manually

If you have trouble using an add_to_platform_X.sh script then you can build the package manually. This is useful if you need to debug why a package is failing to build. If you decide you need to build manually you can do so by launching a singularity shell for the appropriate platform and running the spack install command. To make the command for launching the singularity container with all of the correct volume bindings, admin scripts are available:

  > ssh ifarm1901
  > /scigroup/spack/admin/singshell_centos7.sh
  source /cvmfs/oasis.opensciencegrid.org/jlab/epsci/centos/7.7.1908/share/spack/setup-env.sh
  Singularity>

The last line indicates you are now running in a singularity shell. The line before it that starts with "source /cvmfs/..." SHOULD BE COPIED AND EXECUTED WITHIN THE SHELL. This is important since the setup-env.sh script is not run automatically when the singularity shell is created. Sourcing it is necessary to setup your environment for working with the spack instance.

continuing the example ...

  Singularity> source /cvmfs/oasis.opensciencegrid.org/jlab/epsci/centos/7.7.1908/share/spack/setup-env.sh
  Singularity> spack install -j16 clhep@2.4.1.3 %gcc@9.3.0 target=x86_64 

The above will build the clhep package version 2.4.1.3 using the GCC 9.3.0 compiler and make the binaries for the generic x86_64 target. Note that by default spack will build with microcode for the specific processor in use on the machine on which you are compiling. Adding the target=x86_64 ensures all packages are built the same regardless of specifics of the CPU.

It should also be noted that, if there are dependencies for the package you are building the specific versions should be given in the package specification. The syntax for this is beyond the scope of this document and the spack documentation should be consulted for details.

Updating the buildcache aka mirror

The build cache is only useful if someone is trying to run [#untethered|untethered]. Individual packages (tarballs) can be made from existing spack package builds with the buildcache create command. It is also possible to generate buildcache packages from all packages in a repository or (probably) an Environment with a single command. Look into the spack help for details. Once the buildcache packages are built, you need to rebuild the index. Here are some specific commands:

  1. spack buildcache create -r -a -u -d . zlib%gcc@10.2.1
  2. spack buildcache update-index -k -d /scigroup/spack/mirror

Misc. Notes

Here are some miscellaneous notes to catch useful tidbits not covered in the previous sections.

Mac OS X + Docker

The default disk format for Mac OS X is non-case-sensitive. It automatically translates file and directory names to give the illusion that it is case sensitive. This works fine except when you have two files in the same directory whose name only differs by case. This becomes an issue if you are building spack packages for Linux using Docker and are doing so in a directory from the local disk (bound in the Docker container). I saw this with the ncurses package failing with errors related to E/E_TERM and A/APPLE_TERM (I may not be remembering the exact file names correctly).

One work-around is to create a disk image using Disk Utility and choose the format to be "Mac OS Extended (Case-sensitive, Journaled)". Mount the disk image and bind that to the docker container. This will give you a case sensitive persistent disk (i.e. survives after the container exits).

If you do not care about persistence, then just build in a directory in the Docker container's temporary file system. You can always save to a buildcache from there and copy just the buildcache file out of the container.