GeezerVison

You are right, ↓ benchmark remove -s, slower 0.02us than -s because import

>python -m timeit -n 1000 "import socket" "print(socket.__all__)"
1000 loops, best of 5: 424 usec per loop

>python -m timeit -n 1000 "import socket" "print(socket.__all__)"
1000 loops, best of 5: 371 usec per loop

1.14x faster

The benchmark only imports socket once. the other 999 imports are no-op as socket in in sys.modules.

I carried out two parameter tests, one is with -s, is in the pr description, 1.15x. One is to do without -s in the comments 1.14x

Am I right? I'm not familiar with benchmark, sorry

I'm wrong

You can use hyperfine to measure the entire Python process, in which case using a python -c pass test is useful as a baseline:

❯ hyperfine -w 16 -u microsecond 'python -c pass'
Benchmark 1: python -c pass
  Time (mean ± σ):     10921.2 µs ± 844.5 µs    [User: 8745.1 µs, System: 2091.4 µs]
  Range (min … max):   9720.3 µs … 15591.1 µs    242 runs
❯ hyperfine -w 16 -u microsecond 'python -c "import socket; socket.__all__"'
Benchmark 1: python -c "import socket; socket.__all__"
  Time (mean ± σ):     14554.3 µs ± 1154.4 µs    [User: 12061.3 µs, System: 2368.5 µs]
  Range (min … max):   12826.1 µs … 20450.0 µs    206 runs

Using a PGO+LTO build on macOS, so this will just be measuring the import selectors change, and saves about 1.1 ms:

❯ hyperfine --warmup 16 \
--prepare "git checkout socket"                                   './python.exe -c "import socket; socket.__all__"' \
--prepare "git checkout 6239d41527d5977aa5d44e4b894d719bc045860e" './python.exe -c "import socket;  socket.__all__"'
Benchmark 1: ./python.exe -c "import socket; socket.__all__"
  Time (mean ± σ):      16.2 ms ±   0.8 ms    [User: 13.1 ms, System: 2.5 ms]
  Range (min … max):    15.5 ms …  21.7 ms    91 runs

  Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs.

Benchmark 2: ./python.exe -c "import socket;  socket.__all__"
  Time (mean ± σ):      17.3 ms ±   1.0 ms    [User: 13.9 ms, System: 2.7 ms]
  Range (min … max):    16.6 ms …  25.1 ms    88 runs

  Warning: Statistical outliers were detected. Consider re-running this benchmark on a quiet system without any interferences from other programs.

Summary
  ./python.exe -c "import socket; socket.__all__" ran
    1.06 ± 0.08 times faster than ./python.exe -c "import socket;  socket.__all__"

Also: 1ms Moving the couple of import arrays into two functions that call it:

Can you also make the array module import lazy in this PR?

I measured that the change saves 0.7 ms on import socket: 1.72 ms => 1.24 ms.

timeit:

$ ./python -m timeit -s 'import sys; state=dict(sys.modules)' 'import socket; del socket; sys.modules.clear(); sys.modules.update(state)' 
200 loops, best of 5: 1.72 msec per loop

$ git switch pr/121424 
$ ./python -m timeit -s 'import sys; state=dict(sys.modules)' 'import socket; del socket; sys.modules.clear(); sys.modules.update(state)' 
200 loops, best of 5: 1.24 msec per loop

hyperfine:

$ hyperfine -w 16 './python -c "import socket"'
  Time (mean ± σ):      14.6 ms ±   0.7 ms    [User: 11.9 ms, System: 2.6 ms]
  Range (min … max):    13.4 ms …  16.5 ms    192 runs
 
$ git switch pr/121424 
$ hyperfine -w 16 './python -c "import socket"'
  Time (mean ± σ):      13.9 ms ±   0.6 ms    [User: 11.4 ms, System: 2.4 ms]
  Range (min … max):    12.6 ms …  15.5 ms    197 runs

Thank you! I just replaced the incorrect benchmark results in the PR with the correct ones

Also: 1ms Moving the couple of import arrays into two functions that call it:

Can you also make the array module import lazy in this PR?

@Wulian233 Please could you do this as well?

Of course! I just lazy imported array and haven't done full benchmarking yet. I also mentioned this optimization in 3.14.rst, do you think it should be included? @hugovk

• Improve import time of :mod:socket by lazy importing modules and
  writing :data:!socket.errorTab as a constant, which results in
  a 30% speed-up in the import time pyperformance benchmarks.

What do you think of this👀

For What's New, we can follow the example @AlexWaygood wrote for 3.13, which grouped a few import improvements together:

Several standard library modules have had their import times significantly improved. For example, the import time of the typing module has been reduced by around a third by removing dependencies on re and contextlib. Other modules to enjoy import-time speedups include email.utils, enum, functools, importlib.metadata, and threading. (Contributed by Alex Waygood, Shantanu Jain, Adam Turner, Daniel Hollas, and others in gh-109653.)

Let's do the same with #118761 in 3.14. We have two under that issue so far.

I recommend we also group this PR under #118761 as well -> rename this PR title gh-118761: ....

So we can follow something like that now, or leave it out for now and add a grouped summary later.

Omit details that are not interested to the end user.

I would also remove any mention from Ehat's New -- this is an insignificant change. I am sure that if you measure import time of different modules, you will find larger changes between versions (maybe even between bugfix releases) without anybody noticing.

I recommend we also group this PR under #118761 as well -> rename this PR title gh-118761: ....

Okay, now this pr belongs to 118761. I revert the changes to 3.14 so we can write them together future when there are more modules optimizations :)

Merged. Thanks for your nice enhancement @Wulian233.