This tutorial illustrates how Spack can be used to generate module files for the software that has been installed. Both hierarchical and non-hierarchical deployments will be discussed in details and we will show how to customize the content and naming of each module file.
At the end of the tutorial readers should have a clear understanding of:
- What module files are and how they are used on HPC clusters
- How Spack generates module files for the software it installs
- Which Spack commands can be used to manage module files
- How module files generated by Spack can be customized
and be confident that Spack can deal with all of the common use cases that occur when maintaining software installations on HPC clusters.
To prepare for this tutorial we are going to install a small but representative set of software that includes different configurations of the same packages and some external packages. To keep the installations manageable, let's start by uninstalling everything from earlier in the tutorial:
$ spack uninstall -ay
and by enabling tcl
module files, which are disabled by default since Spack v0.20:
$ spack config add "modules:default:enable:[tcl]"
The first thing that we need is the module tool itself. In the tutorial we will use
lmod
because it can work with both hierarchical and non-hierarchical layouts.
$ spack install lmod
Once the module tool is installed we need to have it available in the
current shell. Installation directories in Spack's store are definitely not easy
to remember, but they can be retrieved with the spack location
command:
$ . $(spack location -i lmod)/lmod/lmod/init/bash
Now we can re-source the setup file and Spack modules will be put in our module path.
$ . spack/share/spack/setup-env.sh
The second step is to build a recent compiler. On first use, Spack
scans the environment and automatically locates the compiler(s)
already available on the system. For this tutorial, however, we want
to use gcc@12.3.0
.
$ spack install gcc@12.3.0
You can get this in your environment using spack load gcc@12.3.0
:
.. literalinclude:: outputs/modules/spack-load-gcc.out :language: console
Now, gcc
is in your PATH
. You can add it to the list of
compilers with spack compiler add
:
.. literalinclude:: outputs/modules/add-compiler.out :language: console
To check which compilers are available you can use spack compiler list
:
.. literalinclude:: outputs/modules/list-compiler.out :language: console
Finally, when you confirmed gcc@12.3.0
is properly registered, clean the environment
with spack unload
:
$ spack unload --all
Finally, we will use Spack to install the packages used in the examples:
$ spack install netlib-scalapack ^openmpi ^openblas
$ spack install netlib-scalapack ^mpich ^openblas
$ spack install netlib-scalapack ^openmpi ^netlib-lapack
$ spack install netlib-scalapack ^mpich ^netlib-lapack
$ spack install py-scipy ^openblas
Module files are an easy way to modify your environment in a controlled
manner during a shell session. In general, they contain the information
needed to run an application or use a library. The module
command is
used to interpret and execute module files. For example, module show
tells you what a module will do when loaded:
.. literalinclude:: outputs/modules/what-are-modules-1.out :language: console
module load
will execute all of the changes shown above:
.. literalinclude:: outputs/modules/what-are-modules-2.out :language: console
and to undo the modifications, you can use module unload
:
.. literalinclude:: outputs/modules/what-are-modules-3.out :language: console
There are two main module systems used in HPC, both installable by Spack.
In this tutorial we will be working with lmod
and be showing examples
with both Tcl and Lua.
This is the original modules tool. It can be installed with Spack using the following command:
$ spack install environment-modules
It was first coded in C in the early 1990s and was later rewritten entirely in Tcl. Long stagnant, the project has been revived in the past few years by Xavier Delaruelle at CEA, and it is now very actively developed. For further details we refer to its documentation.
Lmod is a module system written in Lua, originally created at the "Texas Advanced Computing Center" (TACC) by Robert McLay. You can get it with:
$ spack install lmod
as shown in the :ref:`module_file_tutorial_prerequisites` section. It is a drop-in replacement for Environment Modules, and it works with both Tcl and Lua module files. It is fully compatible with Environment Modules, but it also has many distinguishing features of its own. The main one is the module hierarchy, which simplifies the module UI by only showing modules built with the currently loaded compiler and/or MPI. There are also some unique safety features.
Before we dive into the hands-on sections it's worth explaining how module files are generated by Spack. The following diagram provides a high-level view of the process:
Modules in Spack are generated using configuration files (config.yaml
and modules.yaml
), information from Spack's package recipes, and
Jinja2 templates. Spack comes with Jinja2, an external template engine, so you
do not need to install it yourself.
You may have noticed that we used spack load
in the
:ref:`module_file_tutorial_prerequisites` section above. This is a
built-in mechanism of Spack's -- it's designed so that users on a cluster
or a laptop can quickly get a package into their path, and it understands
Spack's spec syntax. It does not require modules, as Spack needs to
work regardless of whether modules are set up on the system.
As you might expect, you can see what is loaded via spack load
using
spack find
:
.. literalinclude:: outputs/modules/show-loaded.out :language: console
Because Spack is designed to be run on HPC systems, it also generates a module file for every installed package. This allows users unfamiliar with Spack's interface to see things through the module system they're used to. To see this, try:
.. literalinclude:: outputs/modules/module-avail-1.out :language: console
You can module load
any of these. By default, Spack generates modules
named by package-version-compiler-version-hash
, which is a bit hard
to read. We'll show you how to customize this in the following sections.
If you arrived to this point you should have an environment that looks similar to:
.. literalinclude:: outputs/modules/module-avail-2.out :language: console
The non-hierarchical module files that have been generated so far follow
Spack's default rules for module generation.
Taking a look at the gcc
module you'll see, for example:
.. literalinclude:: outputs/modules/module-show-1.out :language: console
As expected, a few environment variables representing paths will be modified by the module file according to the default prefix inspection rules.
Now consider the case that your site has decided that CC
,
CXX
, FC
and F77
modifications should not be
present in module files. What you can do to abide by the rules is to
create a configuration file ${SPACK_ROOT}/etc/spack/modules.yaml
with
the following content:
modules:
default:
tcl:
all:
filter:
exclude_env_vars:
- "CC"
- "CXX"
- "FC"
- "F77"
This can be done either editing the configuration manually, or directly from the command line:
$ spack config add "modules:default:tcl:all:filter:exclude_env_vars:['CC', 'CXX', 'F77', 'FC']"
Next you should regenerate all the module files:
.. literalinclude:: outputs/modules/tcl-refresh-1.out :language: console
If you take a look now at the module for gcc
you'll see that the unwanted
paths have disappeared:
.. literalinclude:: outputs/modules/module-show-2.out :language: console
Another common request at many sites is to avoid exposing software that
is only needed as an intermediate step when building a newer stack.
Let's try to prevent the generation of
module files for anything that is compiled with gcc@11
(the OS provided compiler).
To do this you should add the exclude
keyword to ${SPACK_ROOT}/etc/spack/modules.yaml
:
modules:
default:
tcl:
exclude:
- '%gcc@11'
all:
filter:
exclude_env_vars:
- "CC"
- "CXX"
- "FC"
- "F77"
and regenerate the module files. This time we'll pass the option
--delete-tree
so that Spack will delete the existing module tree and
regenerate a new one, instead of overwriting the files in the existing
directory.
.. literalinclude:: outputs/modules/tcl-refresh-2.out :language: console
.. literalinclude:: outputs/modules/module-avail-3.out :language: console
if you look closely you'll see though that we went too far in
excluding modules: the module for gcc@12.3.0
disappeared as it was
bootstrapped with gcc@11
. To specify exceptions to the exclude
rules you can use include
:
modules:
default:
tcl:
include:
- gcc
exclude:
- '%gcc@11'
all:
filter:
exclude_env_vars:
- "CC"
- "CXX"
- "FC"
- "F77"
include
rules always have precedence over exclude
rules. If you regenerate the modules again:
.. literalinclude:: outputs/modules/tcl-refresh-3.out :language: console
you'll see that now the module for gcc@12.3.0
has reappeared:
.. literalinclude:: outputs/modules/module-avail-4.out :language: console
An additional feature that you can leverage to unclutter the environment is to skip the generation of module files for implicitly installed packages. In this case you only need to add the following line:
modules:
default:
tcl:
exclude_implicits: true
include:
- gcc
exclude:
- '%gcc@11'
all:
filter:
exclude_env_vars:
- "CC"
- "CXX"
- "FC"
- "F77"
to modules.yaml
and regenerate the module file tree as above.
The next step in making module files more user-friendly is to
improve their naming scheme.
To reduce the length of the hash or remove it altogether you can
use the hash_length
keyword in the configuration file:
modules:
default:
tcl:
hash_length: 0
include:
- gcc
exclude:
- '%gcc@11'
all:
filter:
exclude_env_vars:
- "CC"
- "CXX"
- "FC"
- "F77"
If you try to regenerate the module files now you will get an error:
.. literalinclude:: outputs/modules/tcl-refresh-4.out :language: console
Note
We try to check for errors up front!
In Spack we check for errors upfront whenever possible, so don't worry about your module files: as a name clash was detected nothing has been changed on disk.
The problem here is that without the hashes the four different flavors of
netlib-scalapack
map to the same module file name. We can change how
the names are formatted to differentiate them:
modules:
default:
tcl:
hash_length: 0
include:
- gcc
exclude:
- '%gcc@11'
all:
conflict:
- '{name}'
filter:
exclude_env_vars:
- "CC"
- "CXX"
- "FC"
- "F77"
projections:
all: '{name}/{version}-{compiler.name}-{compiler.version}'
netlib-scalapack: '{name}/{version}-{compiler.name}-{compiler.version}-{^lapack.name}-{^mpi.name}'
^python^lapack: '{name}/{version}-{compiler.name}-{compiler.version}-{^lapack.name}'
As you can see it is possible to specify rules that apply only to a
restricted set of packages using anonymous specs
like ^python^lapack
. Here we declare a conflict between any two modules
with the same name, so they cannot be loaded together. We also format the
names of modules according to compiler, compiler version, and MPI provider
name using the spec format syntax.
This allows us to match specs by their dependencies, and format them
based on their DAGs.
.. literalinclude:: outputs/modules/tcl-refresh-5.out :language: console
.. literalinclude:: outputs/modules/module-avail-5.out :language: console
Note
The conflict
directive is Tcl-specific and can't be used in the
lmod
section of the configuration file.
At many sites it is customary to set an environment variable in a
package's module file that points to the folder in which the package
is installed. You can achieve this with Spack by adding an
environment
directive to the configuration file:
modules:
default:
tcl:
hash_length: 0
naming_scheme: '{name}/{version}-{compiler.name}-{compiler.version}'
include:
- gcc
exclude:
- '%gcc@11'
all:
conflict:
- '{name}'
filter:
exclude_env_vars:
- "CC"
- "CXX"
- "FC"
- "F77"
environment:
set:
'{name}_ROOT': '{prefix}'
projections:
all: '{name}/{version}-{compiler.name}-{compiler.version}'
netlib-scalapack: '{name}/{version}-{compiler.name}-{compiler.version}-{^lapack.name}-{^mpi.name}'
^python^lapack: '{name}/{version}-{compiler.name}-{compiler.version}-{^lapack.name}'
Under the hood Spack uses the :meth:`~spack.spec.Spec.format` API to substitute tokens in either environment variable names or values. There are two caveats though:
- The set of allowed tokens in variable names is restricted to
name
,version
,compiler
,compiler.name
,compiler.version
,architecture
- Any token expanded in a variable name is made uppercase, but other than that case sensitivity is preserved
Regenerating the module files results in something like:
.. literalinclude:: outputs/modules/tcl-refresh-6.out :language: console
.. literalinclude:: outputs/modules/module-show-3.out :language: console
As you can see, the gcc
module has the environment variable GCC_ROOT
set.
Sometimes it's also useful to apply environment modifications selectively and target
only certain packages. You can for instance apply modifications to the
openmpi
module as follows:
modules:
default:
tcl:
hash_length: 0
naming_scheme: '{name}/{version}-{compiler.name}-{compiler.version}'
include:
- gcc
exclude:
- '%gcc@11'
all:
conflict:
- '{name}'
filter:
exclude_env_vars:
- "CC"
- "CXX"
- "FC"
- "F77"
environment:
set:
'{name}_ROOT': '{prefix}'
openmpi:
environment:
set:
SLURM_MPI_TYPE: pmi2
OMPI_MCA_btl_openib_warn_default_gid_prefix: '0'
projections:
all: '{name}/{version}-{compiler.name}-{compiler.version}'
netlib-scalapack: '{name}/{version}-{compiler.name}-{compiler.version}-{^lapack.name}-{^mpi.name}'
^python^lapack: '{name}/{version}-{compiler.name}-{compiler.version}-{^lapack.name}'
This time we will be more selective and regenerate only the openmpi
module file:
.. literalinclude:: outputs/modules/tcl-refresh-7.out :language: console
.. literalinclude:: outputs/modules/module-show-4.out :language: console
Spack can also generate module files that contain code to load the
dependencies automatically. You can, for instance generate python
modules that load their dependencies by adding the autoload
directive and assigning it the value direct
:
modules:
default:
tcl:
verbose: true
hash_length: 0
naming_scheme: '{name}/{version}-{compiler.name}-{compiler.version}'
include:
- gcc
exclude:
- '%gcc@11'
all:
conflict:
- '{name}'
filter:
exclude_env_vars:
- "CC"
- "CXX"
- "FC"
- "F77"
environment:
set:
'{name}_ROOT': '{prefix}'
openmpi:
environment:
set:
SLURM_MPI_TYPE: pmi2
OMPI_MCA_btl_openib_warn_default_gid_prefix: '0'
projections:
all: '{name}/{version}-{compiler.name}-{compiler.version}'
netlib-scalapack: '{name}/{version}-{compiler.name}-{compiler.version}-{^lapack.name}-{^mpi.name}'
^python^lapack: '{name}/{version}-{compiler.name}-{compiler.version}-{^lapack.name}'
^python:
autoload: direct
and regenerating the module files for every package that depends on python
:
.. literalinclude:: outputs/modules/tcl-refresh-8.out :language: console
and will contain code to autoload all the dependencies:
.. literalinclude:: outputs/modules/load-direct.out :language: console
In case messages are unwanted during the autoload procedure, it will be
sufficient to omit the line setting verbose: true
in the configuration file above.
So far we worked with non-hierarchical module files, i.e. with module files
that are all generated in the same root directory and don't attempt to
dynamically modify the MODULEPATH
. This results in a flat module structure where
all the software is visible at the same time:
.. literalinclude:: outputs/modules/lmod-intro-avail.out :language: console
This layout is quite simple to deploy, but you can see from the above snippet that nothing prevents users from loading incompatible sets of modules:
.. literalinclude:: outputs/modules/lmod-intro-conflict.out :language: console
Even if conflicts
directives are carefully placed in module files, they:
- won't enforce a consistent environment, but will just report an error
- need constant updates, for instance as soon as a new compiler or MPI library is installed
Hierarchical module files try to
overcome these shortcomings by showing at start-up only a restricted view of what is
available on the system: more specifically only the software that has been installed with
OS provided compilers. Among this software there will be other - usually more recent - compilers
that, once loaded, will prepend new directories to MODULEPATH
unlocking all the software
that was compiled with them. This "unlocking" idea can then be extended arbitrarily to
virtual dependencies, as we'll see in the following section.
The most widely used hierarchy is the so called Core/Compiler/MPI
where, on top
of the compilers, different MPI libraries also unlock software linked to them.
There are just a few steps needed to adapt the modules.yaml
file we used previously:
- enable the
lmod
file generator- change the
tcl
tag tolmod
- remove the
tcl
specificconflict
directive- declare which compilers are considered
core_compilers
- remove the
mpi
related suffixes in projections (as they will be substituted by hierarchies)
After these modifications your configuration file should look like:
modules:
default:
enable::
- lmod
lmod:
core_compilers:
- 'gcc@11'
hierarchy:
- mpi
hash_length: 0
include:
- gcc
exclude:
- '%gcc@11'
all:
filter:
exclude_env_vars:
- "C_INCLUDE_PATH"
- "CPLUS_INCLUDE_PATH"
- "LIBRARY_PATH"
environment:
set:
'{name}_ROOT': '{prefix}'
openmpi:
environment:
set:
SLURM_MPI_TYPE: pmi2
OMPI_MCA_btl_openib_warn_default_gid_prefix: '0'
projections:
all: '{name}/{version}'
^lapack: '{name}/{version}-{^lapack.name}'
Note
- Double colon in configuration files
- The double colon after
enable
is intentional and it serves the purpose of overriding the default list of enabled generators so that onlylmod
will be active (see Overriding entire sections for more details).
The directive core_compilers
accepts a list of compilers. Everything built
using these compilers will create a module in the Core
part of the hierarchy,
which is the entry point for hierarchical module files. It is
common practice to put the OS provided compilers in the list and only build common utilities
and other compilers with them.
If we now regenerate the module files:
.. literalinclude:: outputs/modules/lmod-refresh-1.out :language: console
and update MODULEPATH
to point to the Core
:
$ module purge
$ module unuse $HOME/spack/share/spack/modules/linux-ubuntu18.04-x86_64
$ module use $HOME/spack/share/spack/lmod/linux-ubuntu18.04-x86_64/Core
asking for the available modules will return:
.. literalinclude:: outputs/modules/module-avail-6.out :language: console
Unsurprisingly, the only visible module is gcc
. Loading that we'll unlock
the Compiler
part of the hierarchy:
.. literalinclude:: outputs/modules/module-avail-7.out :language: console
The same holds true also for the MPI
part, that you can enable by loading
either mpich
or openmpi
. Let's start by loading mpich
:
.. literalinclude:: outputs/modules/module-avail-8.out :language: console
.. literalinclude:: outputs/modules/module-load-openblas-scalapack.out :language: console
At this point we can showcase the improved consistency that a hierarchical layout provides over a non-hierarchical one:
.. literalinclude:: outputs/modules/module-swap-mpi.out :language: console
Lmod
took care of swapping the MPI provider for us, and it also substituted the
netlib-scalapack
module to conform to the change in the MPI.
In this way we can't accidentally pull-in two different MPI providers at the
same time or load a module file for a package linked to openmpi
when mpich
is also loaded.
Consistency for compilers and MPI is ensured by the tool.
The hierarchy just shown is already a great improvement over non-hierarchical layouts,
but it still has an asymmetry: LAPACK
providers cover the same semantic role
as MPI
providers, but yet they are not part of the hierarchy.
To be more practical, this means that although we have gained an improved consistency in
our environment when it comes to MPI
, we still have the same problems as we had before
for LAPACK
implementations:
.. literalinclude:: outputs/modules/lapack-conflict.out :language: console
Hierarchies that are deeper than Core
/Compiler
/MPI
are
probably still considered "unusual" or "impractical" at many sites, mainly because
module files are written manually and keeping track of the combinations
among multiple providers quickly becomes quite involved.
For instance, having both MPI
and LAPACK
in the hierarchy
means we must classify software into one of four categories:
- Software that doesn't depend on
MPI
orLAPACK
- Software that depends only on
MPI
- Software that depends only on
LAPACK
- Software that depends on both
to decide when to show it to the user. The situation becomes more involved as the number of virtual dependencies in the hierarchy increases.
We can take advantage of the DAG that Spack maintains for the installed software and solve this combinatorial problem in a clean and automated way. In some sense Spack's ability to manage this combinatorial complexity makes deeper hierarchies feasible.
Coming back to our example, let's add lapack
to the hierarchy and
remove the remaining suffix projection for lapack
:
modules:
default:
enable::
- lmod
lmod:
core_compilers:
- 'gcc@11'
hierarchy:
- mpi
- lapack
hash_length: 0
include:
- gcc
exclude:
- '%gcc@11'
all:
filter:
exclude_env_vars:
- "C_INCLUDE_PATH"
- "CPLUS_INCLUDE_PATH"
- "LIBRARY_PATH"
environment:
set:
'{name}_ROOT': '{prefix}'
openmpi:
environment:
set:
SLURM_MPI_TYPE: pmi2
OMPI_MCA_btl_openib_warn_default_gid_prefix: '0'
projections:
all: '{name}/{version}'
After module files have been regenerated as usual:
.. literalinclude:: outputs/modules/lmod-refresh-2.out :language: console
we can see that now we have additional components in the hierarchy:
.. literalinclude:: outputs/modules/lapack-hier.out :language: console
Both MPI
and LAPACK
providers will now benefit from the same safety features:
.. literalinclude:: outputs/modules/lapack-correct.out :language: console
Because we only compiled py-numpy
with openblas
the module
is made inactive when we switch the LAPACK
provider. The user
environment is now consistent by design!
As briefly mentioned in the introduction, Spack uses Jinja2 to generate each individual module file. This means that you have all of its flexibility and power when it comes to customizing what gets generated!
The templates that Spack uses to generate module files are stored in the
share/spack/templates/module
directory within the Spack prefix, and
they all share the same common structure. Usually, they start with a
header that identifies the type of module being generated. In the case of
hierarchical module files it's:
.. literalinclude:: _spack_root/share/spack/templates/modules/modulefile.lua :language: jinja :lines: 1-6
The statements within double curly brackets {{ ... }}
denote
expressions
that will be evaluated and substituted at module generation time.
The rest of the file is then divided into
blocks
that can be overridden or extended by users, if need be.
Control structures
, delimited by {% ... %}
,
are also permitted in the template language:
.. literalinclude:: _spack_root/share/spack/templates/modules/modulefile.lua :language: jinja :lines: 73-87
The locations where Spack looks for templates are specified
in config.yaml
:
.. literalinclude:: _spack_root/etc/spack/defaults/config.yaml :language: yaml :lines: 32-35
and can be extended by users to employ custom templates, as we'll see next.
Let's assume one of our software is protected by group membership: allowed users belong to the same linux group, and access is granted at group level. Wouldn't it be nice if people that are not yet entitled to use it could receive a helpful message at module load time that tells them who to contact in your organization to be inserted in the group?
To automate the generation of module files with such site-specific behavior
we'll start by extending the list of locations where Spack looks for module
files. Let's create the file ${SPACK_ROOT}/etc/spack/config.yaml
with the content:
config:
template_dirs:
- $HOME/.spack/templates
This tells Spack to also search another location when looking for template files. Next, we need to create our custom template extension in the folder listed above:
{% extends "modules/modulefile.lua" %}
{% block footer %}
-- Access is granted only to specific groups
if not isDir("{{ spec.prefix }}") then
LmodError (
"You don't have the necessary rights to run \"{{ spec.name }}\".\n\n",
"\tPlease write an e-mail to 1234@foo.com if you need further information on how to get access to it.\n"
)
end
{% endblock %}
Let's name this file group-restricted.lua
. The line:
{% extends "modules/modulefile.lua" %}
tells Jinja2 that we are reusing the standard template for hierarchical module files. The section:
{% block footer %}
-- Access is granted only to specific groups
if not isDir("{{ spec.prefix }}") then
LmodError (
"You don't have the necessary rights to run \"{{ spec.name }}\".\n\n",
"\tPlease write an e-mail to 1234@foo.com if you need further information on how to get access to it.\n"
)
end
{% endblock %}
overrides the footer
block.
Finally, we need to add a couple of lines in modules.yaml
to tell Spack which specs
need to use the new custom template. For the sake of illustration let's assume
it's netlib-scalapack
:
modules:
enable::
- lmod
lmod:
core_compilers:
- 'gcc@11'
hierarchy:
- mpi
- lapack
hash_length: 0
include:
- gcc
exclude:
- '%gcc@11'
- readline
all:
filter:
exclude_env_vars:
- "C_INCLUDE_PATH"
- "CPLUS_INCLUDE_PATH"
- "LIBRARY_PATH"
environment:
set:
'{name}_ROOT': '{prefix}'
openmpi:
environment:
set:
SLURM_MPI_TYPE: pmi2
OMPI_MCA_btl_openib_warn_default_gid_prefix: '0'
netlib-scalapack:
template: 'group-restricted.lua'
If we regenerate the module files one last time:
$ spack module lmod refresh -y netlib-scalapack
==> Regenerating lmod module files
we'll find the following at the end of each netlib-scalapack
module file:
-- Access is granted only to specific groups
if not isDir("/home/spack/spack/opt/spack/linux-ubuntu18.04-x86_64/gcc-12.1.0/netlib-scalapack-2.0.2-2p75lzqjbsnev7d2j2osgpkz7ib33oca") then
LmodError (
"You don't have the necessary rights to run \"netlib-scalapack\".\n\n",
"\tPlease write an e-mail to 1234@foo.com if you need further information on how to get access to it.\n"
)
end
and every user that doesn't have access to the software will now be redirected to the right e-mail address where to ask for it!
For future sections of the tutorial, we will not use the gcc@12.3.0
compiler. Since it is currently the default compiler (our current
default is the most recent version of gcc available), we will remove
it now.
$ spack compiler rm gcc@12.3.0
This will ensure the rest of the tutorial goes smoothly for you.