Parse base64-encoded data URIs more efficiently #10434
Open
+111
−13
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
(in some places) Very long data: URIs in source documents are causing outsized memory usage due to various parsing inefficiencies, for instance in Network.URI, TagSoup, and T.P.R.Markdown.source. See e.g. jgm#10075. This change improves the situation in a couple places we can control relatively easily by using an attoparsec text-specialized parser to consume base64-encoded strings. Attoparsec's takeWhile + inClass functions are designed to chew through long strings like this without doing unnecessary allocation, and the improvements in peak heap allocation are significant. One of the observations here is that if you parse something as a valid data: uri it shouldn't need any further escaping so we can short-circuit various processing steps that may unpack/iterate over the chars in the URI.
jgm
reviewed
Dec 4, 2024
Comment on lines
+25
to
+39
| parseBase64String = do | ||
| Sources ((pos, txt):rest) <- getInput | ||
| let r = A.parse pBase64 txt | ||
| case r of | ||
| Done remaining consumed -> do | ||
| let pos' = incSourceColumn pos (T.length consumed) | ||
| setInput $ Sources ((pos', remaining):rest) | ||
| return consumed | ||
| _ -> mzero | ||
|
|
||
| pBase64 :: A.Parser Text | ||
| pBase64 = do | ||
| most <- A.takeWhile1 (A.inClass "A-Za-z0-9+/") | ||
| rest <- A.takeWhile (== '=') | ||
| return $ most <> rest |
There was a problem hiding this comment.
Two thoughts on this:
- Is attoparsec really necessary? Why not just
Data.Text.takeWhile? - My experience is that parsers like this, which just manipulate the input directly using
getInputandsetInput, are problematic in parsec because parsec doesn't realize that input has been consumed. I've had to use a regular parsec parser somewhere in there to make it realize this. One option is just something likecount charsConsumed anyChar, and then you don't need to compute the end position manually...
There was a problem hiding this comment.
- In this spot I can possibly just borrow the fast
inClassfunction from attoparsec and use it with regular texttakeWhile, will have to fiddle with it. - will investigate, I took it for granted that I would make the parsec state happy by fiddling with the input as seen here but that was not based on deep understanding or rigorous analysis.
There was a problem hiding this comment.
It's probably fine to use attoparsec, but there might be a slight speedup if you can avoid it.
On 2: you could try putting this parser under many and see if parsec complains.
jgm
reviewed
Dec 18, 2024
| , textStr ";" <* trace "cool" | ||
| , (mconcat <$> many mediaParam) | ||
| , textStr "base64," <* trace "fine" | ||
| , parseBase64String |
There was a problem hiding this comment.
What if we replace l.1835 with
many1Char (satisfy (A.inClass "A-Za-z0-9+/"))
, manyChar (char '=')
and get rid of the special T.P.Parsing.Base64 module. What is the impact on performance on the sorts of files that were problematic before? I'd like to try the simplest possible thing before worrying about possible optimizations.
jgm
added a commit
that referenced
this pull request
Dec 18, 2024
This patch borrows some code from @silby's PR #10434 and should be regarded as co-authored. This is a lighter-weight patch that only touches the Markdown reader. The basic idea is to speed up parsing of base64 URIs by parsing them with a special path. This should improve the problem noted at #10075. Benchmarks (optimized compilation): Converting the large test.md from #10075 (7.6Mb embedded image) from markdown to json, before: 6182 GCs, 1578M in use, 5.471 MUT (5.316 elapsed), 1.473 GC (1.656 elapsed) after: 951 GCs, 80M in use, .247 MUT (1.205 elapsed), 0.035 GC (0.242 elapsed) For now we leave #10075 open to investigate improvements in HTML rendering with these large data URIs.
|
I borrowed some code from this in a more minimal commit, 1a8da4f |
jgm
added a commit
that referenced
this pull request
Dec 18, 2024
This patch borrows some code from @silby's PR #10434 and should be regarded as co-authored. This is a lighter-weight patch that only touches the Markdown reader. The basic idea is to speed up parsing of base64 URIs by parsing them with a special path. This should improve the problem noted at #10075. Benchmarks (optimized compilation): Converting the large test.md from #10075 (7.6Mb embedded image) from markdown to json, before: 6182 GCs, 1578M in use, 5.471 MUT (5.316 elapsed), 1.473 GC (1.656 elapsed) after: 951 GCs, 80M in use, .247 MUT (1.205 elapsed), 0.035 GC (0.242 elapsed) For now we leave #10075 open to investigate improvements in HTML rendering with these large data URIs. Co-authored-by: Evan Silberman <[email protected]>
jgm
added a commit
that referenced
this pull request
Dec 18, 2024
This patch borrows some code from @silby's PR #10434 and should be regarded as co-authored. This is a lighter-weight patch that only touches the Markdown reader. The basic idea is to speed up parsing of base64 URIs by parsing them with a special path. This should improve the problem noted at #10075. Benchmarks (optimized compilation): Converting the large test.md from #10075 (7.6Mb embedded image) from markdown to json, before: 6182 GCs, 1578M in use, 5.471 MUT, 1.473 GC after: 951 GCs, 80M in use, .247 MUT, 0.035 GC For now we leave #10075 open to investigate improvements in HTML rendering with these large data URIs. Co-authored-by: Evan Silberman <[email protected]>
jgm
added a commit
that referenced
this pull request
Dec 19, 2024
Text.Pandoc.URI: export `pBase64DataURI`. Modify `isURI` to use this and avoid calling network-uri's inefficient `parseURI` for data URIs. Markdown reader: use T.P.URI's `pBase64DataURI` in parsing data URIs. Partially addresses #10075. Obsoletes #10434 (borrowing most of its ideas). Co-authored-by: Evan Silberman <[email protected]>
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.
(in some places)
Very long data: URIs in source documents are causing outsized memory usage due to various parsing inefficiencies, for instance in Network.URI, TagSoup, and T.P.R.Markdown.source. See e.g. #10075.
This change improves the situation in a couple places we can control relatively easily by using an attoparsec text-specialized parser to consume base64-encoded strings. Attoparsec's takeWhile + inClass functions are designed to chew through long strings like this without doing unnecessary allocation, and the improvements in peak heap allocation are significant.
One of the observations here is that if you parse something as a valid data: uri it shouldn't need any further escaping so we can short-circuit various processing steps that may unpack/iterate over the chars in the URI.
The code here is organized a little bit randomly and I expect it needs more work to go in.
Good improvements though. The peak heap allocation for html -> markdown for the example provided by @ebeigarts in #10075 goes down from ~2900MB to ~2500 MB on my computer, and markdown -> json goes from 1577 MB to 73 MB! As discussed in the issue comments HTML reading has, at least, a TagSoup issue.