Improve async runtime scaling #946
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Minoru
approved these changes
Aug 18, 2022
There was a problem hiding this comment.
Hey, sorry it took me so long to take a look at this! The code looks good.
I tested it on my site (https://github.com/Minoru/blog.debiania.in.ua), which only has 187 Markdown files to compile, but also embeds 32 thousand MathJax files (processed with copyFileCompiler). The performance is indeed better, but the scaling is still bad. Furthermore, the sweet spot is now at 2 cores rather than 4 cores (for an 4-core CPU with hyperthreading), which is odd.
$ hyperfine --prepare './debiania-old-1645d9c clean' './debiania-old-1645d9c build'
Benchmark 1: ./debiania-old-1645d9c build
Time (mean ± σ): 15.999 s ± 0.324 s [User: 52.089 s, System: 14.393 s]
Range (min … max): 15.472 s … 16.603 s 10 runs
$ hyperfine --parameter-scan threads 1 9 --prepare './debiania-old-1645d9c clean' './debiania-old-1645d9c build +RTS -N{threads}'
Benchmark 1: ./debiania-old-1645d9c build +RTS -N1
Time (mean ± σ): 13.335 s ± 0.150 s [User: 13.083 s, System: 0.768 s]
Range (min … max): 13.101 s … 13.601 s 10 runs
Benchmark 2: ./debiania-old-1645d9c build +RTS -N2
Time (mean ± σ): 11.098 s ± 0.344 s [User: 15.222 s, System: 2.124 s]
Range (min … max): 10.690 s … 11.691 s 10 runs
Benchmark 3: ./debiania-old-1645d9c build +RTS -N3
Time (mean ± σ): 10.812 s ± 0.601 s [User: 18.158 s, System: 3.344 s]
Range (min … max): 9.991 s … 11.861 s 10 runs
Benchmark 4: ./debiania-old-1645d9c build +RTS -N4
Time (mean ± σ): 11.622 s ± 0.496 s [User: 22.643 s, System: 5.129 s]
Range (min … max): 11.041 s … 12.421 s 10 runs
Benchmark 5: ./debiania-old-1645d9c build +RTS -N5
Time (mean ± σ): 12.484 s ± 0.622 s [User: 28.233 s, System: 7.239 s]
Range (min … max): 11.621 s … 13.492 s 10 runs
Benchmark 6: ./debiania-old-1645d9c build +RTS -N6
Time (mean ± σ): 13.577 s ± 0.473 s [User: 33.883 s, System: 9.680 s]
Range (min … max): 12.752 s … 14.310 s 10 runs
Benchmark 7: ./debiania-old-1645d9c build +RTS -N7
Time (mean ± σ): 14.613 s ± 0.323 s [User: 40.769 s, System: 12.303 s]
Range (min … max): 13.957 s … 14.992 s 10 runs
Benchmark 8: ./debiania-old-1645d9c build +RTS -N8
Time (mean ± σ): 15.961 s ± 0.250 s [User: 51.374 s, System: 14.443 s]
Range (min … max): 15.504 s … 16.296 s 10 runs
Benchmark 9: ./debiania-old-1645d9c build +RTS -N9
Time (mean ± σ): 17.072 s ± 0.263 s [User: 57.513 s, System: 16.108 s]
Range (min … max): 16.535 s … 17.337 s 10 runs
Summary
'./debiania-old-1645d9c build +RTS -N3' ran
1.03 ± 0.07 times faster than './debiania-old-1645d9c build +RTS -N2'
1.07 ± 0.08 times faster than './debiania-old-1645d9c build +RTS -N4'
1.15 ± 0.09 times faster than './debiania-old-1645d9c build +RTS -N5'
1.23 ± 0.07 times faster than './debiania-old-1645d9c build +RTS -N1'
1.26 ± 0.08 times faster than './debiania-old-1645d9c build +RTS -N6'
1.35 ± 0.08 times faster than './debiania-old-1645d9c build +RTS -N7'
1.48 ± 0.09 times faster than './debiania-old-1645d9c build +RTS -N8'
1.58 ± 0.09 times faster than './debiania-old-1645d9c build +RTS -N9'
$ hyperfine --prepare './debiania-new-3a54b21 clean' './debiania-new-3a54b21 build'
Benchmark 1: ./debiania-new-3a54b21 build
Time (mean ± σ): 14.586 s ± 0.188 s [User: 46.536 s, System: 11.439 s]
Range (min … max): 14.244 s … 14.874 s 10 runs
$ hyperfine --parameter-scan threads 1 9 --prepare './debiania-new-3a54b21 clean' './debiania-new-3a54b21 build +RTS -N{threads}'
Benchmark 1: ./debiania-new-3a54b21 build +RTS -N1
Time (mean ± σ): 9.744 s ± 0.020 s [User: 9.390 s, System: 0.611 s]
Range (min … max): 9.703 s … 9.764 s 10 runs
Benchmark 2: ./debiania-new-3a54b21 build +RTS -N2
Time (mean ± σ): 8.385 s ± 0.061 s [User: 11.360 s, System: 1.233 s]
Range (min … max): 8.304 s … 8.494 s 10 runs
Benchmark 3: ./debiania-new-3a54b21 build +RTS -N3
Time (mean ± σ): 8.974 s ± 0.420 s [User: 14.677 s, System: 2.336 s]
Range (min … max): 8.464 s … 9.780 s 10 runs
Benchmark 4: ./debiania-new-3a54b21 build +RTS -N4
Time (mean ± σ): 9.638 s ± 0.633 s [User: 18.245 s, System: 3.308 s]
Range (min … max): 8.705 s … 10.665 s 10 runs
Benchmark 5: ./debiania-new-3a54b21 build +RTS -N5
Time (mean ± σ): 10.708 s ± 0.595 s [User: 23.503 s, System: 5.118 s]
Range (min … max): 9.555 s … 11.704 s 10 runs
Benchmark 6: ./debiania-new-3a54b21 build +RTS -N6
Time (mean ± σ): 11.364 s ± 0.385 s [User: 28.192 s, System: 6.938 s]
Range (min … max): 10.809 s … 11.875 s 10 runs
Benchmark 7: ./debiania-new-3a54b21 build +RTS -N7
Time (mean ± σ): 13.280 s ± 0.382 s [User: 35.940 s, System: 9.331 s]
Range (min … max): 12.556 s … 13.666 s 10 runs
Benchmark 8: ./debiania-new-3a54b21 build +RTS -N8
Time (mean ± σ): 14.559 s ± 0.205 s [User: 46.252 s, System: 11.489 s]
Range (min … max): 14.222 s … 14.912 s 10 runs
Benchmark 9: ./debiania-new-3a54b21 build +RTS -N9
Time (mean ± σ): 15.182 s ± 0.164 s [User: 49.571 s, System: 12.024 s]
Range (min … max): 14.943 s … 15.395 s 10 runs
Summary
'./debiania-new-3a54b21 build +RTS -N2' ran
1.07 ± 0.05 times faster than './debiania-new-3a54b21 build +RTS -N3'
1.15 ± 0.08 times faster than './debiania-new-3a54b21 build +RTS -N4'
1.16 ± 0.01 times faster than './debiania-new-3a54b21 build +RTS -N1'
1.28 ± 0.07 times faster than './debiania-new-3a54b21 build +RTS -N5'
1.36 ± 0.05 times faster than './debiania-new-3a54b21 build +RTS -N6'
1.58 ± 0.05 times faster than './debiania-new-3a54b21 build +RTS -N7'
1.74 ± 0.03 times faster than './debiania-new-3a54b21 build +RTS -N8'
1.81 ± 0.02 times faster than './debiania-new-3a54b21 build +RTS -N9'
Could it be that with copyFIleCompiler the individual compiler invocations is so short that the contention on IORef is killing the gains? I guess it could be ameliorated if threads took multiple jobs from the queue at once, but that complicates the design. Perhaps I should write a recursiveCopyFilesCompier that will copy the entirety of MathJax in one go :)
|
Even though the performance gains aren't as great as I hoped, it drastically improves resource usage (open file handles), as discussed in haskellfoundation/error-message-index#444, so I will merge this in and push a release. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This noticeably improves performance on most sites I tried, though I usually need to copy
one of the blogposts around a 1000 times for the effect to become more noticeable and
measurable.
Still in draft because the error you get on cyclic dependencies got much worse, that'sfixable though.