caseSensitiveFS=false (the default) scan performance penalty scales up, and it can be extreme

imsodin · April 22, 2024, 8:28am

If I remember right, CI produces apks from PRs. So you can update the syncthing submodule with that patch and PR that to the android repo. Otherwise of you have docker, building is very simple.

calmh · April 22, 2024, 8:41am

The listdir cache is intended to prevent that behaviour, but I suspect the cache time is too low. I think I mentioned that as some point.

bt90 · April 1, 2025, 2:52pm

Sorry for necromancing this thread, but I finally managed to resurrect my old PR:

github.com/syncthing/syncthing

chore(fs): speed up case normalization

main ← ascii_fastpath

opened 10:48AM - 01 Apr 25 UTC

bt90

+77 -21

### Purpose Resurrecting https://github.com/syncthin …### Testing fs │ ../old.txt │ ../new.txt │ │ sec/op │ sec/op vs base │ 43.68n ± 4% 12.19n ± 1% -72.09% (p=0.000 n=10) 200.65n ± 2% 59.49n ± 3% -70.35% (p=0.000 n=10) 95.50n ± 2% 59.10n ± 2% -38.12% (p=0.000 n=10) 122.5n ± 1% 131.4n ± 1% +7.27% (p=0.000 n=10) 339.9n ± 2% 309.2n ± 1% -9.05% (p=0.000 n=10) 183.6n ± 2% 174.3n ± 1% -5.04% (p=0.000 n=10) 456.6n ± 1% 440.5n ± 1% -3.53% (p=0.000 n=10) 625.9n ± 1% 595.6n ± 1% -4.83% (p=0.000 n=10) 516.2n ± 1% 495.5n ± 1% -4.02% (p=0.000 n=10) 214.1n 150.4n -29.72% │ ../old.txt │ ../new.txt │ │ B/op │ B/op vs base │ 0.000 ± 0% 0.000 ± 0% ~ (p=1.000 n=10) ¹ 24.00 ± 0% 24.00 ± 0% ~ (p=1.000 n=10) ¹ 24.00 ± 0% 24.00 ± 0% ~ (p=1.000 n=10) ¹ 0.000 ± 0% 0.000 ± 0% ~ (p=1.000 n=10) ¹ 32.00 ± 0% 32.00 ± 0% ~ (p=1.000 n=10) ¹ 32.00 ± 0% 32.00 ± 0% ~ (p=1.000 n=10) ¹ 0.000 ± 0% 0.000 ± 0% ~ (p=1.000 n=10) ¹ 48.00 ± 0% 48.00 ± 0% ~ (p=1.000 n=10) ¹ 48.00 ± 0% 48.00 ± 0% ~ (p=1.000 n=10) ¹ ² +0.00% ² geomean │ ../old.txt │ ../new.txt │ │ allocs/op │ allocs/op vs base │ 0.000 ± 0% 0.000 ± 0% ~ (p=1.000 n=10) ¹ 1.000 ± 0% 1.000 ± 0% ~ (p=1.000 n=10) ¹ 1.000 ± 0% 1.000 ± 0% ~ (p=1.000 n=10) ¹ 0.000 ± 0% 0.000 ± 0% ~ (p=1.000 n=10) ¹ 1.000 ± 0% 1.000 ± 0% ~ (p=1.000 n=10) ¹ 1.000 ± 0% 1.000 ± 0% ~ (p=1.000 n=10) ¹ 0.000 ± 0% 0.000 ± 0% ~ (p=1.000 n=10) ¹ 1.000 ± 0% 1.000 ± 0% ~ (p=1.000 n=10) ¹ 1.000 ± 0% 1.000 ± 0% ~ (p=1.000 n=10) ¹ ² +0.00% ² geomean Latin1 testcase is easily outweighed by the ASCII and unicode improvements :slightly_smiling_face:

Feel free to test the current main branch to see if it has a meaningful impact for case-insensitive scanning