sys.settrace dramatic slowdown in 3.12

[nedbat]

Bug report

Checklist

• I am confident this is a bug in CPython, not a bug in a third-party project
• I have searched the CPython issue tracker, and am confident this bug has not been reported before

A clear and concise description of the bug

A bug report in coverage (nedbat/coveragepy#1665) is reduced here to a pure do-nothing trace function for sys.settrace.

Reproduction:

% git clone https://github.com/nedbat/ndindex

% cd ndindex

% python3.11 -VV
Python 3.11.4 (main, Jun  7 2023, 08:42:37) [Clang 14.0.3 (clang-1403.0.22.14.1)]

% python3.11 -m venv v311

% ./v311/bin/pip install numpy

% python3.12 -VV
Python 3.12.0b4 (main, Jul 11 2023, 20:38:26) [Clang 14.0.3 (clang-1403.0.22.14.1)]

% python3.12 -m venv v312

% ./v312/bin/pip install --pre --extra-index https://pypi.anaconda.org/scientific-python-nightly-wheels/simple numpy

% # Run the code without the trace function.
% for x in 1 2 3; time ./v311/bin/python -m ndindex.tests.justdoit notrace
./v311/bin/python -m ndindex.tests.justdoit notrace  7.05s user 0.37s system 108% cpu 6.813 total
./v311/bin/python -m ndindex.tests.justdoit notrace  7.04s user 0.34s system 109% cpu 6.724 total
./v311/bin/python -m ndindex.tests.justdoit notrace  7.00s user 0.35s system 109% cpu 6.696 total

% # 3.12 is slightly faster without the trace function.
% for x in 1 2 3; time ./v312/bin/python -m ndindex.tests.justdoit notrace
./v312/bin/python -m ndindex.tests.justdoit notrace  6.56s user 0.30s system 106% cpu 6.422 total
./v312/bin/python -m ndindex.tests.justdoit notrace  6.48s user 0.31s system 106% cpu 6.359 total
./v312/bin/python -m ndindex.tests.justdoit notrace  6.38s user 0.28s system 107% cpu 6.217 total

% # 3.11 with tracing is about 3x slower.
% for x in 1 2 3; time ./v311/bin/python -m ndindex.tests.justdoit trace
./v311/bin/python -m ndindex.tests.justdoit trace  22.64s user 0.51s system 101% cpu 22.772 total
./v311/bin/python -m ndindex.tests.justdoit trace  21.92s user 0.46s system 101% cpu 21.979 total
./v311/bin/python -m ndindex.tests.justdoit trace  21.55s user 0.35s system 102% cpu 21.379 total

% # 3.12 with tracing is 7x slower.
% for x in 1 2 3; time ./v312/bin/python -m ndindex.tests.justdoit trace
./v312/bin/python -m ndindex.tests.justdoit trace  49.47s user 0.40s system 100% cpu 49.676 total
./v312/bin/python -m ndindex.tests.justdoit trace  49.53s user 0.39s system 100% cpu 49.784 total
./v312/bin/python -m ndindex.tests.justdoit trace  50.44s user 0.38s system 100% cpu 50.739 total

I don't know if the unusual numpy build has something to do with this.

Linked PRs

• GH-107674: Avoid allocating boxed ints for sys.settrace line events #107780
• [3.12] GH-107674: Avoid allocating boxed ints for sys.settrace line events (GH-107780) #107841
• gh-107674: Improve performance of sys.settrace #114986
• gh-107674: Improve performance of sys.settrace #117133
• gh-107674: Remove some unnecessary code in instrumentation code #117393
• gh-107674: Lazy load line number to improve performance of tracing #118127

[neonene]

This issue can be seen starting from the commit 411b169 with usual numpy-1.25.2.

[AlexWaygood]

cc. @markshannon

[nedbat]

@markshannon Is there any news about this?

[nedbat]

@iritkatriel perhaps you have some idea?

[iritkatriel]

@iritkatriel perhaps you have some idea?

Did #107780 help?

[gaogaotiantian]

Hi @nedbat , I circled back to this issue. If you are still interested, could you try the patch of #114986 and see if the result is improved?

[nedbat]

@gaogaotiantian Thanks. I tried #114986, and it didn't run faster. I used the https://github.com/Quansight-Labs/ndindex repo test suite:
3.11: 160s
3.12: 255s
3.13 main: 248s
pr114986: 245s

I wish I had better news.

[gaogaotiantian]

Oh, did you test it with an empty trace function?

[nedbat]

I will do more tests, but the ultimate goal is to perform better in real-world scenarios. Here is more data from a subset of the test suite (-k test_chunking), which looks much more promising. It's hard to get consistent numbers, but this looks good. I would merge this change.

Python version	no tracing	Python trace	C trace
3.11.8	1.19s	7.39s	2.46s
3.12.2	1.12s	11.57s	3.71s
3.13.0a3+ heads/main:a95b1a56bb	1.14s	10.37s	3.89s
3.13.0a3+ heads/pr/114986:cb09c55758	1.14s	6.48s	2.91s

[gaogaotiantian]

Thank you for the test. Yes the PR only improves the performance of the actual trace function - which is normally the major cost for tracing tools. Good to know it helps with the more real-life benchmark. Unfortunately, due to the change of the tracing mechanism, the overhead for sys.settrace will always be larger than the old way - I did not see a quick solution to that.

[markshannon]

Now that the eval_breaker is part of the thread state, here's an alternative approach that might work.

tstate->tracing only needs one bit, so use the low bit in the version part of the eval breaker.
For non-tracing functions we expect to always see the low bit set to 0.
For tracing functions we expect to always see the low bit set to 1.

If the code and global versions differ do the following:

• If tracing bit is 0, update the version number (setting the low bit to 0), and instrument.
• If the tracing bit is 1:
  • If the version number matches except the low bit: do nothing
  • Otherwise: update the version number (setting the low bit to 1), and clear any instrumentation.

I think this is correct, but I've not drawn up a full state diagram.

[encukou]

#114986 caused refleaks tests (./python -m test -R 3:3 test_frame) to fail.
If not fixed in 24 hours, I'll revert the commit to unblock buildbots. (I hope I'll get have time to investigate & come up with a proper fix, but can't promise that.)

[gaogaotiantian]

@encukou See response in #116098

[gaogaotiantian]

Hi @nedbat , do you have a chance to test the performance of coverage on 3.13b against 3.12? I made a few optimizations and I'd hope the performance issue is relieved a bit.

[nedbat]

I'm having a hard time re-creating the scenarios I used before :(

[gaogaotiantian]

That's okay. Does coveragepy have nightly validations for 3.13b? Is there a basic timing benchmark that we can refer to? Just a ballpark would be helpful.

[asmeurer]

You can try running the ndindex tests, which was the original thing I noticed this issue with in nedbat/coveragepy#1665. Just clone

https://github.com/Quansight-Labs/ndindex

then install the packages in requirements-dev.txt and run

time pytest

(coverage is enabled automatically). On my computer with Python 3.11.9, the tests take 2:48 and with Python 3.12.0 they take 4:21 (both with coverage 7.5.2). If you want to run the tests with coverage disabled you should remove the coverage options from pytest.ini. Note that there is a degree of variation in the test times due to the use of hypothesis. You might want to set the --hypothesis-seed to avoid this.

[nedbat]

Coverage.py has an overnight job that runs its test suite on nightly builds of Python, including 3.13 and 3.14 now: https://github.com/nedbat/coveragepy/actions/runs/9265205433 There isn't a timing test in there though.

[gaogaotiantian]

Okay thanks @nedbat and @asmeurer . I was hoping for some quick and existing benchmark but it's okay that we don't have any for now. Just wondering if we should close this issue now that we have implemented some optimizations.

[nedbat]

I have a benchmarking framework, but it's fragile. I will try to patch it up later this week.

[nedbat]

I ran benchmarks again on 3.11, 3.12, and 3.13:

cov	proj	python3.11	python3.12	python3.13	3.12 vs 3.11	3.13 vs 3.12
761	mashumaro	56.4s	62.6s	67.9s	111%	108%
761	pygments	25.1s	56.2s	38.9s	224%	69%
761	mypy	109.4s	113.8s	104.4s	104%	92%

The specific versions were the newly released 3.11.10, 3.12.6, and 3.13.0rc2. The test suites were run three times, and the median time used.

[gaogaotiantian]

Thanks @nedbat . I'm a bit surprised that on some cases 3.13 is slower than 3.12 - I did not think of anything that we did to slow down the tracing. Glad to see we had some significant improvements on certain cases.

[nedbat]

I continue to get people reporting serious performance problems on 3.12. Is there any chance of an improvement?

[junkmd]

Considering the issue observed in enthought/comtypes#775, where execution time increased nearly tenfold, there is a performance "regression" in Python 3.12.

Only one bugfix release remains for Python 3.12:
https://pythonpeps.python.org/pep-0693/

If #107841 is not merged by then, I believe the Python community will miss an opportunity to reconsider this issue.

If "the fix patch" is merged before Python 3.12 transitions to security-fix only status, I would accept enabling coverage measurement in the comtypes CI tests for all Python versions (and waiting until 3.12.10 is released). Otherwise, only Python 3.12 will be tested without coverage measurement in the comtypes CI pipeline.

As discussed in #107841, I understand that this issue doesn't involve "broken" programs but rather a "performance improvement", and that the changes are complex to accept during the bugfix phase. However, fixing this could potentially benefit projects and the wider ecosystem running on Python 3.12.

I would like to, once again, hear the opinions of the core developers and the community on this matter.

[gaogaotiantian]

Ah, @markshannon merged #128350 in main but did not backport it to 3.12 and 3.13 due to conflicts. #107841 is not the problem fixer. I believe backporting #128350 should fix the issue in 3.12 and 3.13.

[junkmd]

I tested whether #107841 is the problem fixer.

I built Python from v3.12.8 and also from the 3.12 branch with 37d8b90 cherry-picked.
And I compared the execution time of comtypes test cases.

with cov	37d8b90 changes	Python 3.12.8
yes	1101.106s	1254.337s
no	57.869s	46.481s

The build with 37d8b90 changes is about 153s faster than v3.12.8.
However, this performance improvement is still not significant.

I haven't tested backporting #128350 yet because I'm not sure how to resolve conflicts.