- This reduces concretization time for r-rminer from over 1 minute to
only 16 seconds.
- OrderedDict ends up taking a *lot* of time to compare larger specs.
- An OrderedDict isn't actually needed here. It's actually not possible
to find duplicates, and we end up sorting the contents of the
OrderedDict anyway.
- This is an optimization to the way traverse_edges iterates over
successors.
- Previous version called dependencies_dict(), which involved a lot of
redundant work (creating dicts and calling caonical_deptype)
- Spack ends up constructing compilers frequently from YAML data.
- This caches the result of parsing the compiler config
- The logic in compilers/__init__.py could use a bigger cleanup, but this
makes concretization much faster for now.
- on macOS, this also ensures that xcrun is called only twice, as opposed
to every time a new compiler object is constructed.
This adds a workflow section on how to use spack on Docker.
It provides an example on the best-practices I collected over the
last months and circumvents the common pitfalls I tapped in.
Works with MPI, CUDA, Modules, execution as root, etc.
Background: Developed initially for PIConGPU.
* Make --trusted default when running spack gpg list
Currently running `spack gpg list` with no arguments returns nothing. You must supply either the `--trusted` or the `--signing` options. The idea here is to return some initial data to the user when the command is run. The alternative is to return an error, telling the user to select one of the two options.
* Add an extra test case for the empty list command
Fixes the issue with code coverage
This isn't a rework of the concretizer but it speeds things up a LOT.
The main culprits were:
1. Variant code, `provider_index`, and `concretize.py` were calling
`spec.package` when they could use `spec.package_class`
- `spec.package` looks up a package instance by `Spec`, which requires a
(fast-ish but not that fast) DAG compare.
- `spec.package_class` just looks up the package's class by name, and you
should use this when all you need is metadata (most of the time).
- not really clear that the current way packages are looked up is
necessary -- we can consider refactoring that in the future.
2. `Repository.repo_for_pkg` parses a `str` argument into a `Spec` when
called with one, via `@_autospec`, but this is not needed.
- Add some faster code to handle strings directly and avoid parsing
This speeds up concretization 3-9x in my limited tests. Still not super
fast but much more bearable:
Before:
- `spack spec xsdk` took 33.6s
- `spack spec dealii` took 1m39s
After:
- `spack spec xsdk` takes 6.8s
- `spack spec dealii` takes 10.8s
* Do not call setup_package for fake installs
- setup package could fail if ``setup_dependent_environment`` or other
routines expected to use executables from dependencies
- xpetsc and boost try to get python config variables in
`setup_dependent_package`; this would cause them not to be
fake-installable
* Remove vestigial deptype_query argument to Spec.traverse()
- The `deptype_query` argument isn't used anymore -- it's only passed
around and causes confusion when calling traverse.
- Get rid of it and just keep the `deptypes` argument
* Don't print redundant messages when installing dependencies
- `do_install()` was originally depth-first recursive, and printed "<pkg>
already installed in ..." multiple times for packages as recursive
calls encountered them.
- For much cleaner output, use spec.traverse(order='post') to install
dependencies instead
