After this PR, from sys.gettotalrefcount: 17647 to sys.gettotalrefcount: 14775.

Can you please try to measure the important on performance of this change? You can use:

Note: Please use a release build (./configure && make).

Can you please try to measure the important on performance of this change? You can use:

Note: Please use a release build (./configure && make).

Hi, victor. Got this benchmark result.

$ python3 -m pyperf command -- ./python -S -c pass
.....................
WARNING: the benchmark result may be unstable

• the standard deviation (1.62 ms) is 11% of the mean (14.3 ms)

Try to rerun the benchmark with more runs, values and/or loops.
Run 'python3 -m pyperf system tune' command to reduce the system jitter.
Use pyperf stats, pyperf dump and pyperf hist to analyze results.
Use --quiet option to hide these warnings.

command: Mean +- std dev: 14.3 ms +- 1.6 ms

$ python3 -m pyperf command -- ./python -S -c pass --duplicate=100
.....................
WARNING: the benchmark result may be unstable

• the standard deviation (2.33 ms) is 15% of the mean (15.1 ms)

Try to rerun the benchmark with more runs, values and/or loops.
Run 'python3 -m pyperf system tune' command to reduce the system jitter.
Use pyperf stats, pyperf dump and pyperf hist to analyze results.
Use --quiet option to hide these warnings.

command: Mean +- std dev: 15.1 ms +- 2.3 ms

Hi, victor. Got this benchmark result.

What did you measure? What is the reference (unmodified master branch) and what is the modified code (with your PR)?

You can store the benchmark result in a file (-o bench.json) and then compare two files (python3 -m pyperf compare_to ref.json pr.json).

What is the dictionary size of the interned strings at Python exit (in unicode_release_interned())?

Please rebase your PR on master, I modified unicodeobject.c to make singletons per interpreter.

Hi, victor. Got this benchmark result.

What did you measure? What is the reference (unmodified master branch) and what is the modified code (with your PR)?

You can store the benchmark result in a file (-o bench.json) and then compare two files (python3 -m pyperf compare_to ref.json pr.json).

python3 is master, ./python with tihs PR. I test it again after I fix the assert problem.

python3 is master, ./python with tihs PR

You tested ./python is both cases. Usually, to measure the performance of a change, you need a reference (your "python3") and the modified Python (your "./python").

python3 is master, ./python with tihs PR

You tested ./python is both cases. Usually, to measure the performance of a change, you need a reference (your "python3") and the modified Python (your "./python").

ok, Look like I confused test python3's performance with use pyperf of python3 to compare other python version's load performance.
you mean this one, right?:

python3 -m pyperf command ./python_ref  xxx -o ref.json
python3 -m pyperf command ./python_pr xxx -o pr.json
python3 -m pyperf compare_to ref.json pr.json

python3 -m pyperf command ./python_ref  xxx -o ref.json
python3 -m pyperf command ./python_pr xxx -o pr.json
python3 -m pyperf compare_to ref.json pr.json

Yeah, that would work :-)

Anyway, I expect no significant difference. So the most important data point is the number of items in the dictionary at exit.

python3 -m pyperf command ./python_ref  xxx -o ref.json
python3 -m pyperf command ./python_pr xxx -o pr.json
python3 -m pyperf compare_to ref.json pr.json

Yeah, that would work :-)

Anyway, I expect no significant difference. So the most important data point is the number of items in the dictionary at exit.

Got this result ;) :
$ python3.9 -m pyperf compare_to ref.json pr.json
Benchmark hidden because not significant (1): command

What is the dictionary size of the interned strings at Python exit (in unicode_release_interned())?

(gdb) p PyDict_Size(interned)
$1 = 1471

I wrote PR #21269 which clears interned strings in the correct order.

I merged my PR #21269.