Overcoming Software

Haddock Performance

I was recently made aware that haddock hasn’t been working, at all, on the Mecury code base. I decided to investigate. Watching htop, haddock slowly accumulated memory, until it exploded in use and invoked the OOM killer.

My laptop has 64GB of RAM. What.

I rebooted, and tried again. With no other programs running, haddock was able to complete. I enabled -v2 and --optghc=-ddump-timings, which printed out GHC timing information and Haddock time/memory information. With these flags, I could see that HTML generation alone was allocating 800GB of RAM.

I decided to look at the source code and see if there were any low hanging fruit. Fortunately, there was!

Don’t use WriterT for logging

This section culminated in this PR #1543, which I’ll reference. At time of writing, it has been merged.

The first thing that jumped out at me is that haddock used WriterT for logging. Even worse, it used WriterT [String]. This is maybe the slowest possible logging system you can imagine.

At least, use the CPS Writer

WriterT has a big problem with space leaks. Even the strict WriterT has this issue. The only WriterT that can avoid it is the CPS variant, newly available in mtl-2.3 in Control.Monad.Writer.CPS. This is documented in Infinite Negative Utility’s post “Writer Monads and Space Leaks”, which references two posts from Gabriella Gonzalez to the mailing list in 2012 and 2013.

Don’t use [String] or String

Beyond just leaking space, the String format for log messages is extremely inefficient. This is equal to a [Char], which builds up as a big thunk in memory until the whole WriterT computation can complete. Each Char takes 2 words of memory, and a [Char] takes (1 + 3N) words + 2N where N is the number of characters. Or, ~5 words per character. On a 64 bit machine, each word is 8 bytes, so each character costs 40 bytes. A UTF-8 encoded ByteString will use 1 to 4 bytes per character. Using a ByteString would make the representation much more compact, but these things get concatenated a bunch, and a Builder is the appropriate choice for an O(1) append.

Switching to CPS.WriterT [Builder] instead of WriterT [String] helps, but we’re not done yet. [] is a bad choice for a WriterT, since tell will mappend the lists. mappend on lists can result in bad performance if it isn’t associated correctly - (((a ++ b) ++ c) ++ d) ++ e is accidentally quadratic, since we’ll traverse over the a list for every single ++ call. A “difference list” has much faster appends, since it can associate things correctly regardless of how you construct it.

To make it easier to use the API, I created an ErrorMessages type:

newtype ErrorMessages = ErrorMessages { unErrorMessages :: [Builder] -> [Builder] }
    deriving newtype (Semigroup, Monoid)

runErrMsgM :: ErrMsgM a -> (a, ErrorMessages)
runErrMsgM = runWriter . unErrMsgM

singleMessage :: Builder -> ErrorMessages
singleMessage m = ErrorMessages $ (m :)

errorMessagesToList :: ErrorMessages -> [Builder]
errorMessagesToList messages = unErrorMessages messages []

Don’t use nub

There were a few places that called nub. nub is O(n^2) on a linked list - it must search the entire list for every element in the list to verify uniqueness.

That last one is especially gnarly. We’re doing an O(n^2) job, leaking space along the way, and finally we accumulate the big list- only to do an O(n^2) traversal over it to delete duplicates.

Fortunately, each call site of nub can be replaced with the easy ordNub:

ordNub :: Ord a => [a] -> [a]
ordNub = Set.toList . Set.fromList

This also sorts the list, which may not be desired. A more cumbersome implementation does this:

ordNub :: Ord a => [a] -> [a]
ordNub = go Set.empty
  where
    go seen [] = []
    go seen (x:xs) 
        | Set.member x seen = 
            go seen xs
        | otherwise = 
            x : go (Set.insert x seen) xs

Results

This small change resulted in a huge improvement in performance for my test case. Running haddock on the persistent-test library, I observed a 30% improvement in the time to generate documenation with total memory use 4% better. Allocations went from 42GB to 25GB.

I didn’t bother profiling to determine this as a hot-spot - it’s always wrong to use WriterT as a logger. A further performance improvement would be to remove WriterT entirely and simply output the messages directly. Then instead of retaining a big difference list of log messages, you can just print them right then and there.

xhtml and [String]

This section is represented by this haddock PR and this xhtml PR.

haddock uses a library xhtml for generating the HTML. This library is old - the initial copyright is 1999. xhtml predates ByteString entirely, which has an earliest copyright of 2003.

Anyway, we have a similar problem. The Html type is defined like this:

newtype Html = Html { getHtmlElements :: [HtmlElement] }

data HtmlElement
      = HtmlString String
      | HtmlTag {
              markupTag      :: String,
              markupAttrs    :: [HtmlAttr],
              markupContent  :: Html
              }

-- | Attributes with name and value.
data HtmlAttr = HtmlAttr String String

The xhtml library uses ++ on lists all over the place. The renderHtml' function uses ShowS, fortunately - this is difference list of Char, so we probably won’t be seeing pessimal performance.

Like the above PR to remove WriterT [String] and replace it with a difference list of Builder, I did that to xhtml. All explicit lists are now difference lists, and all String are replaced with Builder.

The performance results are impressive:

  Haddock Head xhtml Builder Absolute Difference Relative Change
HTML allocations 1134 MB 1141 MB +7 MB 0.6% worse
HTML time: 380 ms 198 ms -182 ms 47.9% improvement
Total Memory: 554 MB 466 MB -88 MB 15.9% improvement
Total Allocated: 16.0 GB 16.0 GB 0 No change
Max residency: 238 MB 195 MB -43 MB 18.1% improvement
Total Time: 10.88 s 6.526s s -4.354 s 40% improvement

Avoiding [] and String halves the time to render HTML, and results in a 40% overall improvement in the time to run haddock. While we don’t allocate any fewer memory during HTML generation, we’re using 16% less total memory and maximum residency is down by 18%.

Conclusion

Haskell performance doesn’t have to be hard. If you avoid common footguns like WriterT, [], String, nub, etc. then your code will probably be pretty quick.

Picking the low hanging fruit is usually worthwhile, even if you haven’t spent the effort determining the real problem. Profiling shows that haddock spends an enormous amount of time generating object code - a necessary step for any module that has TemplateHaskell enabled. With GHC 9.6, we’ll be able to pass -fprefer-byte-code, which will use the much faster byte code representation instead of object code.

Even in HTML generation, profiling indicates that we spend the majority of time doing fixChar - the process of escaping a character into an HTML appropriate Builder. We also spend a bunch of time regenerating HTML for re-exports - the HTML documentation for a datatype, function, type class, etc is generated fresh for every module that exports it.

Even if HTML were perfectly optimized, Haddock’s current design creates a huge [Interface], where each Interface is a module that you are generating documentation for. This [Interface] must be retained in memory, because it is passed to each “component” of the documentation build. Refactoring haddock to stream these interfaces isn’t obvious, since some doc building steps require summary of the entire [Interface] in order to proceed.

Figuring out a fix for the “real problems” would have been much more difficult than these easy fixes, which have still made a huge difference in overall perforamnce.