So after having squeezed the last bit of performance out of some Haskell I am using to break tweet data into n-grams, I'm running up against a space leak problem. When I profile, the GC uses about 60-70% of the process and there are significant memory portions dedicated to drag. Hopefully, some Haskell guru will be able to suggest when I'm going wrong.

{-# LANGUAGE OverloadedStrings, BangPatterns #-}
import Data.Maybe
import qualified Data.ByteString.Char8 as B
import qualified Data.HashMap.Strict as H
import Text.Regex.Posix
import Data.List
import qualified Data.Char as C

isClassChar a = C.isAlphaNum a || a == ' ' || a == '\'' ||
                a == '-' || a == '#' || a == '@' || a == '%'

cullWord :: B.ByteString -> B.ByteString
cullWord w = B.map C.toLower $ B.filter isClassChar w

procTextN :: Int -> B.ByteString -> [([B.ByteString],Int)]
procTextN n t = H.toList $ foldl' ngram H.empty lines
  where !lines        = B.lines $ cullWord t
        ngram tr line = snd $ foldl' breakdown (base,tr) (B.split ' ' line)
        base          = replicate (n-1) ""

breakdown :: ([B.ByteString], H.HashMap [B.ByteString] Int) -> B.ByteString -> ([B.ByteString],H.HashMap [B.ByteString] Int)
breakdown (st@(s:ss),tree) word =
  newStack `seq` expandedWord `seq` (newStack,expandedWord)
      where newStack     = ss ++ [word]
            expandedWord = updateWord (st ++ [word]) tree

updateWord :: [B.ByteString] -> H.HashMap [B.ByteString] Int -> H.HashMap [B.ByteString] Int
updateWord w h = H.insertWith (+) w 1 h

main = do
  test2 <- B.readFile "canewobble"
  print $ filter (\(a,b) -> b > 100) $ 
     sortBy (\(a,b) (c,d) -> compare d b) $ procTextN 3 test2
haskell memory-leaks space
Edit
Report