reorder packages.yaml: requirements first, then preferences (#33741)

* reorder packages.yaml: requirements first, then preferences
* expand preferences vs reuse vs requirements
This commit is contained in:
Harmen Stoppels 2022-11-08 01:16:11 +01:00 committed by GitHub
parent b3b675157c
commit 3693622edf
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -302,88 +302,31 @@ microarchitectures considered during the solve are constrained to be compatible
host Spack is currently running on. For instance, if this option is set to ``true``, a
user cannot concretize for ``target=icelake`` while running on an Haswell node.
.. _package-preferences:
-------------------
Package Preferences
-------------------
Spack can be configured to prefer certain compilers, package
versions, dependencies, and variants during concretization.
The preferred configuration can be controlled via the
``~/.spack/packages.yaml`` file for user configurations, or the
``etc/spack/packages.yaml`` site configuration.
Here's an example ``packages.yaml`` file that sets preferred packages:
.. code-block:: yaml
packages:
opencv:
compiler: [gcc@4.9]
variants: +debug
gperftools:
version: [2.2, 2.4, 2.3]
all:
compiler: [gcc@4.4.7, 'gcc@4.6:', intel, clang, pgi]
target: [sandybridge]
providers:
mpi: [mvapich2, mpich, openmpi]
At a high level, this example is specifying how packages should be
concretized. The opencv package should prefer using GCC 4.9 and
be built with debug options. The gperftools package should prefer version
2.2 over 2.4. Every package on the system should prefer mvapich2 for
its MPI and GCC 4.4.7 (except for opencv, which overrides this by preferring GCC 4.9).
These options are used to fill in implicit defaults. Any of them can be overwritten
on the command line if explicitly requested.
Each ``packages.yaml`` file begins with the string ``packages:`` and
package names are specified on the next level. The special string ``all``
applies settings to *all* packages. Underneath each package name is one
or more components: ``compiler``, ``variants``, ``version``,
``providers``, and ``target``. Each component has an ordered list of
spec ``constraints``, with earlier entries in the list being preferred
over later entries.
Sometimes a package installation may have constraints that forbid
the first concretization rule, in which case Spack will use the first
legal concretization rule. Going back to the example, if a user
requests gperftools 2.3 or later, then Spack will install version 2.4
as the 2.4 version of gperftools is preferred over 2.3.
An explicit concretization rule in the preferred section will always
take preference over unlisted concretizations. In the above example,
xlc isn't listed in the compiler list. Every listed compiler from
gcc to pgi will thus be preferred over the xlc compiler.
The syntax for the ``provider`` section differs slightly from other
concretization rules. A provider lists a value that packages may
``depend_on`` (e.g, MPI) and a list of rules for fulfilling that
dependency.
.. _package-requirements:
--------------------
Package Requirements
--------------------
You can use the configuration to force the concretizer to choose
specific properties for packages when building them. Like preferences,
these are only applied when the package is required by some other
request (e.g. if the package is needed as a dependency of a
request to ``spack install``).
Spack can be configured to always use certain compilers, package
versions, and variants during concretization through package
requirements.
An example of where this is useful is if you have a package that
is normally built as a dependency but only under certain circumstances
(e.g. only when a variant on a dependent is active): you can make
sure that it always builds the way you want it to; this distinguishes
package configuration requirements from constraints that you add to
``spack install`` or to environments (in those cases, the associated
packages are always built).
Package requirements are useful when you find yourself repeatedly
specifying the same constraints on the command line, and wish that
Spack respects these constraints whether you mention them explicitly
or not. Another use case is specifying constraints that should apply
to all root specs in an environment, without having to repeat the
constraint everywhere.
The following is an example of how to enforce package properties in
``packages.yaml``:
Apart from that, requirements config is more flexible than constraints
on the command line, because it can specify constraints on packages
*when they occur* as a dependency. In contrast, on the command line it
is not possible to specify constraints on dependencies while also keeping
those dependencies optional.
The package requirements configuration is specified in ``packages.yaml``
keyed by package name:
.. code-block:: yaml
@ -452,15 +395,15 @@ under ``all`` are disregarded. For example, with a configuration like this:
cmake:
require: '%gcc'
Spack requires ``cmake`` to use ``gcc`` and all other nodes (including cmake dependencies)
to use ``clang``.
Spack requires ``cmake`` to use ``gcc`` and all other nodes (including ``cmake``
dependencies) to use ``clang``.
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Setting requirements on virtual specs
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
A requirement on a virtual spec applies whenever that virtual is present in the DAG. This
can be useful for fixing which virtual provider you want to use:
A requirement on a virtual spec applies whenever that virtual is present in the DAG.
This can be useful for fixing which virtual provider you want to use:
.. code-block:: yaml
@ -470,8 +413,8 @@ can be useful for fixing which virtual provider you want to use:
With the configuration above the only allowed ``mpi`` provider is ``mvapich2 %gcc``.
Requirements on the virtual spec and on the specific provider are both applied, if present. For
instance with a configuration like:
Requirements on the virtual spec and on the specific provider are both applied, if
present. For instance with a configuration like:
.. code-block:: yaml
@ -483,6 +426,66 @@ instance with a configuration like:
you will use ``mvapich2~cuda %gcc`` as an ``mpi`` provider.
.. _package-preferences:
-------------------
Package Preferences
-------------------
In some cases package requirements can be too strong, and package
preferences are the better option. Package preferences do not impose
constraints on packages for particular versions or variants values,
they rather only set defaults -- the concretizer is free to change
them if it must due to other constraints. Also note that package
preferences are of lower priority than reuse of already installed
packages.
Here's an example ``packages.yaml`` file that sets preferred packages:
.. code-block:: yaml
packages:
opencv:
compiler: [gcc@4.9]
variants: +debug
gperftools:
version: [2.2, 2.4, 2.3]
all:
compiler: [gcc@4.4.7, 'gcc@4.6:', intel, clang, pgi]
target: [sandybridge]
providers:
mpi: [mvapich2, mpich, openmpi]
At a high level, this example is specifying how packages are preferably
concretized. The opencv package should prefer using GCC 4.9 and
be built with debug options. The gperftools package should prefer version
2.2 over 2.4. Every package on the system should prefer mvapich2 for
its MPI and GCC 4.4.7 (except for opencv, which overrides this by preferring GCC 4.9).
These options are used to fill in implicit defaults. Any of them can be overwritten
on the command line if explicitly requested.
Package preferences accept the follow keys or components under
the specific package (or ``all``) section: ``compiler``, ``variants``,
``version``, ``providers``, and ``target``. Each component has an
ordered list of spec ``constraints``, with earlier entries in the
list being preferred over later entries.
Sometimes a package installation may have constraints that forbid
the first concretization rule, in which case Spack will use the first
legal concretization rule. Going back to the example, if a user
requests gperftools 2.3 or later, then Spack will install version 2.4
as the 2.4 version of gperftools is preferred over 2.3.
An explicit concretization rule in the preferred section will always
take preference over unlisted concretizations. In the above example,
xlc isn't listed in the compiler list. Every listed compiler from
gcc to pgi will thus be preferred over the xlc compiler.
The syntax for the ``provider`` section differs slightly from other
concretization rules. A provider lists a value that packages may
``depends_on`` (e.g, MPI) and a list of rules for fulfilling that
dependency.
.. _package_permissions:
-------------------