I took a close look at how Debian builds the Python interpreter,
because I noticed it ran substantially faster than the one in nixpkgs
and I was curious why.
One thing that I found made a material difference in performance was
this pair of linker flags (passed to the compiler):
-Wl,-O1 -Wl,-Bsymbolic-functions
In other words, effectively the linker gets passed the flags:
-O1 -Bsymbolic-functions
Doing the same thing in nixpkgs turns out to make the interpreter
run about 6% faster, which is quite a big win for such an easy
change. So, let's apply it.
---
I had not known there was a `-O1` flag for the *linker*!
But indeed there is.
These flags are unrelated to "link-time optimization" (LTO), despite
the latter's name. LTO means doing classic compiler optimizations
on the actual code, at the linking step when it becomes possible to
do them with cross-object-file information. These two flags, by
contrast, cause the linker to make certain optimizations within the
scope of its job as the linker.
Documentation is here, though sparse:
https://sourceware.org/binutils/docs-2.31/ld/Options.html
The meaning of -O1 was explained in more detail in this LWN article:
https://lwn.net/Articles/192624/
Apparently it makes the resulting symbol table use a bigger hash
table, so the load factor is smaller and lookups are faster. Cool.
As for -Bsymbolic-functions, the documentation indicates that it's a
way of saving lookups through the symbol table entirely. There can
apparently be situations where it changes the behavior of a program,
specifically if the program relies on linker tricks to provide
customization features:
https://bugs.launchpad.net/ubuntu/+source/xfe/+bug/644645https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=637184#35
But I'm pretty sure CPython doesn't permit that kind of trick: you
don't load a shared object that tries to redefine some symbol found
in the interpreter core.
The stronger reason I'm confident using -Bsymbolic-functions is
safe, though, is empirical. Both Debian and Ubuntu have been
shipping a Python built this way since forever -- it was introduced
for the Python 2.4 and 2.5 in Ubuntu "hardy", and Debian "lenny",
released in 2008 and 2009. In those 12 years they haven't seen a
need to drop this flag; and I've been unable to locate any reports
of trouble related to it, either on the Web in general or on the
Debian bug tracker. (There are reports of a handful of other
programs breaking with it, but not Python/CPython.) So that seems
like about as thorough testing as one could hope for.
---
As for the performance impact: I ran CPython upstream's preferred
benchmark suite, "pyperformance", in the same way as described in
the previous commit. On top of that commit's change, the results
across the 60 benchmarks in the suite are:
The median is 6% faster.
The middle half (aka interquartile range) is from 4% to 8% faster.
Out of 60 benchmarks, 3 come out slower, by 1-4%. At the other end,
5 are at least 10% faster, and one is 17% faster.
So, that's quite a material speedup! I don't know how big the
effect of these flags is for other software; but certainly CPython
tends to do plenty of dynamic linking, as that's how it loads
extension modules, which are ubiquitous in the stdlib as well as
popular third-party libraries. So perhaps that helps explain why
optimizing the dynamic linker has such an impact.
Without this flag, the configure script prints a warning at the end,
like this (reformatted):
If you want a release build with all stable optimizations active
(PGO, etc), please run ./configure --enable-optimizations
We're doing a build to distribute to people for day-to-day use,
doing things other than developing the Python interpreter. So
that's certainly a release build -- we're the target audience for
this recommendation.
---
And, trying it out, upstream isn't kidding! I ran the standard
benchmark suite that the CPython developers use for performance
work, "pyperformance". Following its usage instructions:
https://pyperformance.readthedocs.io/usage.html
I ran the whole suite, like so:
$ nix-shell -p ./result."$variant" --run '
cd $(mktemp -d); python -m venv venv; . venv/bin/activate
pip install pyperformance
pyperformance run -o ~/tmp/result.'"$variant"'.json
'
and then examined the results with commands like:
$ python -m pyperf compare_to --table -G \
~/tmp/result.{$before,$after}.json
Across all the benchmarks in the suite, the median speedup was 16%.
(Meaning 1.16x faster; 14% less time).
The middle half of them ranged from a 13% to a 22% speedup.
Each of the 60 benchmarks in the suite got faster, by speedups
ranging from 3% to 53%.
---
One reason this isn't just the default to begin with is that, until
recently, it made the build a lot slower. What it does is turn on
profile-guided optimization, which means first build for profiling,
then run some task to get a profile, then build again using the
profile. And, short of further customization, the task it would use
would be nearly the full test suite, which includes a lot of
expensive and slow tests, and can easily take half an hour to run.
Happily, in 2019 an upstream developer did the work to carefully
select a more appropriate set of tests to use for the profile:
https://github.com/python/cpython/commit/4e16a4a31https://bugs.python.org/issue36044
This suite takes just 2 minutes to run. And the resulting final
build is actually slightly faster than with the much longer suite,
at least as measured by those standard "pyperformance" benchmarks.
That work went into the 3.8 release, but the same list works great
if used on older releases too.
So, start passing that --enable-optimizations flag; and backport
that good-for-PGO set of tests, so that we use it on all releases.
The ./configure script prints a warning when passed this flag,
starting with 3.7:
configure: WARNING: unrecognized options: --with-threads
The reason is that there's no longer such a thing as a build
without threads.
Eliminate the warning, by only passing the flag on the older releases
that accept it.
Upstream change and discussion:
https://github.com/python/cpython/commit/a6a4dc816https://bugs.python.org/issue31370
- Replaced python override from the final stdenv, instead we
propagate our bootstrap python to stage4 and override both
CF and xnu to use it.
- Removed CF argument from python interpreters, this is redundant
since it's not overidden anymore.
- Inherit CF from stage4, making it the same as the stdenv.
When `makeWrapperArgs` variable is not set, `declare -p makeWrapperArgs`
will return with 1 and print an error message to stderr.
I did not handle the non-existence case in b0633406cb
because I thought `mk-python-derivation` will always define `makeWrapperArgs`
but `wrapProgram` can be called independently. And even with `mk-python-derivation`,
`makeWrappers` will not be set unless explicitly declared in the derivation
because of https://github.com/NixOS/nix/issues/1461.
I was lead to believe that because the builds were succeeding and I confirmed
that the mechanism fails when the variable is not defined and `-o nounset` is enabled.
It appears that `wrapPython` setup hook is not running under `-o nounset`, though,
invaldating the assumption.
Now we are checking that the variable exists before checking its type, which
will get rid of the warning and also prevent future error when `-o nounset`
is enabled in the setup hook.
For more information, see the discussion at
https://github.com/NixOS/nixpkgs/commit/a6bb2ede232940a96150da7207a3ecd15eb6328
Bash takes an assignment of a string to an array variable:
local -a user_args
user_args="(foo bar)"
to mean appending the string to the array, not parsing the string into
an array as is the case when on the same line as the declaration:
local -a user_args="(foo bar)"
b0633406cb extracted the declaration before
the newly branched code block, causing string makeWrapperArgs being added
to the array verbatim.
Since local is function scoped, it does not matter if we move it inside
each of the branches so we fix it this way.
When `makeWrapperArgs` is a Bash array, we only passed the first
item to `wrapProgram`. We need to use `"${makeWrapperArgs[@]}"`
to extract all the items. But that breaks the common string case so
we need to handle that case separately.
This will turn manylinux support back on by default.
PIP will now do runtime checks against the compatible glibc version to
determine if the current interpreter is compatible with a given
manylinux specification. However it will not check if any of the
required libraries are present.
The motivation here is that we want to support building python packages
with wheels that require manylinux support. There is no real change for
users of source builds as they are still buildings packages from source.
The real noticeable(?) change is that impure usages (e.g. running `pip
install package`) will install manylinux packages that previously
refused to install.
Previously we did claim that we were not compatible with manylinux and
thus they wouldn't be installed at all.
Now impure users will have basically the same situation as before: If
you require some wheel only package it didn't work before and will not
properly work now. Now the program will fail during runtime vs during
installation time.
I think it is a reasonable trade-off since it allows us to install
manylinux packages with nix expressions and enables tools like
poetry2nix.
This should be a net win for users as it allows wheels, that we
previously couldn't really support, to be used.
It's a year until the final release but this will give a chance to test
out certain features and how it integrates with other packages.
https://www.python.org/dev/peps/pep-0596/
There were no new changes in version 3.5.9; 3.5.9 was released only because of a CDN caching problem,
which resulted in some users downloading a prerelease version of the 3.5.8 .xz source tarball.
Apart from the version number, 3.5.9 is identical to the proper 3.5.8 release.
