A benchmark of builder->lazy bytestring versus builder->streaming bytestring (originally it had a different purpose). See results below.

benchcsv.hs

{-# LANGUAGE OverloadedStrings, PackageImports #-}
-- |
-- Copyright   : (c) 2010-2011 Simon Meier
-- License     : BSD3-style (see LICENSE)
--
-- Maintainer  : Simon Meier <[email protected]>
-- Stability   : experimental
-- Portability : tested on GHC only
--
-- Running example for documentation of Data.ByteString.Lazy.Builder
--
-- module Main (main) where

-- **************************************************************************
-- CamHac 2011: An introduction to Data.ByteString.Lazy.Builder
-- **************************************************************************

------------------------------------------------------------------------------
-- Benchmarking
------------------------------------------------------------------------------

main :: IO ()
main = do
    putStrLn "Encoding the maxiTable"
    putStrLn $ "Total length in bytes: " ++
        (show $ L.length $ encodeUtf8CSV maxiTable)
    putStrLn $ "Chunk lengths: " ++
        (show $ map S.length $ L.toChunks $ encodeUtf8CSV maxiTable)
    putStrLn ""
    defaultMain
      [ benchNF
      , benchBuilderUtf8     -- LB.length $ toLazyByteString builder
      , benchBuilderUtf8Streaming -- Q.length $ toStreamingByteString builder
      , benchBuilderEncodingUtf8  -- LB.length $ toLazyByteString builder
      , benchBuilderEncodingUtf8Streaming  -- Q.length $ toStreamingByteString builder
      
      ]
  where
    encodeUtf8CSV = B.toLazyByteString . renderTableBE

{- The Encoding Problem
 ----------------------

 Encoding: Conversion from a Haskell value to a sequence of bytes.


 Efficient encoding implementation:

   1. represent sequence of bytes as a list of byte arrays (chunks)
   2. generate chunks that are large on average
   3. avoid intermediate copies/datastructures

 Compositionality:

   4. support fast append


 Problem: Provide a library for defining compositional, efficient encodings.

-}



{- Data.ByteString.Lazy.Builder
 ------------------------------

 A solution to the "Encoding Problem"  (based on the code of blaze-builder).

 Builder creation:

   word8   :: Word8 -> Builder
   int64LE :: Int64 -> Builder
   floatBE :: Float -> Builder
   ....


 Builder composition via its Monoid instance:

   word8 10 `mappend` floatBE 1.4


 Builder execution by converting it to a lazy bytestring:

   toLazyByteString :: Builder -> L.ByteString

-}


{- Typical users of Builders
 ---------------------------

 binary, text, aeson, blaze-html, blaze-textual, warp, snap-server, ...

 => they want support for maximal performance!
 => use of Builders is rather local: in rendering/encoding functions.

-}



{- Notable properties
 --------------------

 * Built-in UTF-8 support: very hard to get efficient otherwise.

     stringUtf8 :: String -> Builder
     intDec :: Int -> Builder
     intHex :: Int -> Builder

 * Fine-grained control over when to copy/reference existing bytestrings

 * EDSL for defining low-level Encodings of bounded values (e.g., Int, Char)
   to improve speed of escaping and similar operations.

 * If used together with iteratee-style IO: no 'unsafePerformIO' required

-}


{- An example problem:
 ---------------------

 Rendering a table in comma-separated-value (CSV) format using UTF-8 encoded
 Unicode characters.

 * We are willing to fuse table-rendering with UTF8-encoding to achieve better
   performance.

-}


import qualified Data.ByteString                     as S
import qualified Data.ByteString.Lazy                as L

import           Data.ByteString.Lazy.Builder                         as B
import           Data.ByteString.Lazy.Builder.ASCII                   as B

import Data.Monoid
import Data.Foldable (foldMap)

import Criterion.Main
import Control.DeepSeq


-- To be used in a later optimization
import           Data.ByteString.Builder.Prim ( (>*<), (>$<) )
import qualified Data.ByteString.Builder.Prim as E

-- To be used in a later comparison
import qualified Data.DList                                      as D
import qualified Codec.Binary.UTF8.Light                         as Utf8Light
import qualified Data.String.UTF8                                as Utf8String
import qualified Data.Text.Lazy                                  as TL
import qualified Data.Text.Lazy.Encoding                         as TL
import qualified Data.Text.Lazy.Builder                          as TB
import qualified Data.Text.Lazy.Builder.Int                      as TB

import Data.Char (ord)
import qualified Data.Binary.Builder                             as BinB
import Data.Functor.Identity
import qualified Data.ByteString.Streaming as Streaming 
------------------------------------------------------------------------------
-- Simplife CSV Tables
------------------------------------------------------------------------------

data Cell = StringC String
          | IntC Int
          deriving( Eq, Ord, Show )

type Row   = [Cell]
type Table = [Row]

-- Example data
strings :: [String]
strings =  ["hello", "\"1\"", "λ-wörld"]

table :: Table
table = [map StringC strings, map IntC [-3..3]]


-- | The rendered 'table':
--
-- > "hello","\"1\"","λ-wörld"
-- > -3,-2,-1,0,1,2,3
--


-- | A bigger table for benchmarking our encoding functions.
maxiTable :: Table
maxiTable = take 5000 $ cycle table


------------------------------------------------------------------------------
-- String based rendering
------------------------------------------------------------------------------

renderString :: String -> String
renderString cs = "\"" ++ concatMap escape cs ++ "\""
  where
    escape '\\' = "\\"
    escape '\"' = "\\\""
    escape c    = return c

renderCell :: Cell -> String
renderCell (StringC cs) = renderString cs
renderCell (IntC i)     = show i

renderRow :: Row -> String
renderRow []     = ""
renderRow (c:cs) = renderCell c ++ concat [',' : renderCell c' | c' <- cs]

renderTable :: Table -> String
renderTable rs = concat [renderRow r ++ "\n" | r <- rs]

-- 1.36 ms
benchString :: Benchmark
benchString = bench "renderTable maxiTable" $ nf renderTable maxiTable

-- 1.36 ms
benchStringUtf8 :: Benchmark
benchStringUtf8 = bench "utf8 + renderTable maxiTable" $
  nf (L.length . B.toLazyByteString . B.stringUtf8 . renderTable) maxiTable


-- using difference lists:  0.91 ms
--
--  (++) is a performance-grinch!


------------------------------------------------------------------------------
-- Builder based rendering
------------------------------------------------------------------------------

-- better syntax for `mappend`
-- infixr 4 <>
-- (<>) :: Monoid m => m -> m -> m
-- (<>) = mappend

-- As a reminder:
--
-- import  Data.ByteString.Lazy.Builder       as B
-- import  Data.ByteString.Lazy.Builder.Utf8  as B

renderStringB :: String -> Builder
renderStringB cs = B.charUtf8 '"' <> foldMap escape cs <> B.charUtf8 '"'
  where
    escape '\\' = B.charUtf8 '\\' <> B.charUtf8 '\\'
    escape '\"' = B.charUtf8 '\\' <> B.charUtf8 '"'
    escape c    = B.charUtf8 c

renderCellB :: Cell -> Builder
renderCellB (StringC cs) = renderStringB cs
renderCellB (IntC i)     = B.intDec i

renderRowB :: Row -> Builder
renderRowB []     = mempty
renderRowB (c:cs) =
    renderCellB c <> mconcat [ B.charUtf8 ',' <> renderCellB c' | c' <- cs ]

renderTableB :: Table -> Builder
renderTableB rs = mconcat [renderRowB r <> B.charUtf8 '\n' | r <- rs]

-- 0.81ms
benchBuilderUtf8 :: Benchmark
benchBuilderUtf8 = bench "utf8 + renderTableB maxiTable + lazy" $
  nf (L.length . B.toLazyByteString . renderTableB) maxiTable


benchBuilderUtf8Streaming :: Benchmark
benchBuilderUtf8Streaming = bench "utf8 + renderTableB maxiTable  + streaming " $
    nfIO ((Streaming.length . Streaming.toStreamingByteString . renderTableB) maxiTable)

-- 1.11x  faster than DList

-- However: touching the whole table 'nf maxiTable' takes  0.27ms

-- 1.16x  faster than DList on the code path other than touching all data
--        (0.91 - 0.27) / (0.82 - 0.27)


------------------------------------------------------------------------------
-- Baseline: Touching all data
------------------------------------------------------------------------------

instance NFData Cell where
  rnf (StringC cs) = rnf cs
  rnf (IntC i)     = rnf i

-- 0.27 ms
benchNF :: Benchmark
benchNF = bench "nf maxiTable" $ nf id maxiTable


------------------------------------------------------------------------------
-- Exploiting bounded encodings
------------------------------------------------------------------------------

{- Why 'Bounded Encodings'?
 --------------------------

 Hot code of encoding implementations:

 * Appending Builders: Optimized already.

 * Encoding primitive Haskell values: room for optimization:

     - reduce buffer-free checks
     - remove jumps/function calls
     - hoist constant values out of inner-loops
       (e.g., the loop for encoding the elements of a list)

 * Bounded encoding:
     an encoding that never takes more than a fixed number of bytes.

     - intuitively: (Int,     Ptr Word8 -> IO (Ptr Word8))
                     ^bound   ^ low-level encoding function

     - compositional: coalesce buffer-checks, ...

       E.encodeIfB :: (a -> Bool)
                   -> BoundedEncoding a -> BoundedEncoding a -> BoundedEncoding a
       E.charUtf8  :: BoundedEncoding Char
       (>*<)       :: BoundedEncoding a -> BoundedEncoding b -> BoundedEncoding (a, b)

       (>$<)       :: (b -> a) -> BoundedEncoding a -> BoundedEncoding b

       ^ BoundedEncodings are contrafunctors; like most data-sinks


     - Implementation relies heavily on inlining to compute bounds and
       low-level encoding code during compilation.
-}
renderStringBE :: String -> Builder
renderStringBE cs = B.charUtf8 '"' <> foldMap escape cs <> B.charUtf8 '"'
  where
    escape '\\' = B.charUtf8 '\\' <> B.charUtf8 '\\'
    escape '\"' = B.charUtf8 '\\' <> B.charUtf8 '\"'
    escape c    = B.charUtf8 c

renderCellBE :: Cell -> Builder
renderCellBE (StringC cs) = renderStringBE cs
renderCellBE (IntC i)     = B.intDec i

renderRowBE :: Row -> Builder
renderRowBE []     = mempty
renderRowBE (c:cs) =
    renderCellBE c <> mconcat [ B.charUtf8 ',' <> renderCellBE c' | c' <- cs ]

renderTableBE :: Table -> Builder
renderTableBE rs = mconcat [renderRowBE r <> B.charUtf8 '\n' | r <- rs]

-- 0.65 ms
benchBuilderEncodingUtf8 :: Benchmark
benchBuilderEncodingUtf8 = bench "utf8 + renderTableBE maxiTable + lazy" $
  nf (L.length . B.toLazyByteString . renderTableBE) maxiTable

benchBuilderEncodingUtf8Streaming :: Benchmark
benchBuilderEncodingUtf8Streaming = bench "utf8 + renderTableBE maxiTable  + streaming " $
        nfIO $ (Streaming.length . Streaming.toStreamingByteString . renderTableBE) maxiTable



-- 1.4x faster than DList based

-- 1.7x faster than DList based on code other than touching all data


------------------------------------------------------------------------------
-- Difference-list based rendering
------------------------------------------------------------------------------

type DString = D.DList Char

renderStringD :: String -> DString
renderStringD cs = return '"' <> foldMap escape cs <> return '"'
  where
    escape '\\' = D.fromList "\\\\"
    escape '\"' = D.fromList "\\\""
    escape c    = return c

renderCellD :: Cell -> DString
renderCellD (StringC cs) = renderStringD cs
renderCellD (IntC i)     = D.fromList $ show i

renderRowD :: Row -> DString
renderRowD []     = mempty
renderRowD (c:cs) =
    renderCellD c <> mconcat [ return ',' <> renderCellD c' | c' <- cs ]

renderTableD :: Table -> DString
renderTableD rs = mconcat [renderRowD r <> return '\n' | r <- rs]

-- 0.91 ms
benchDListUtf8 :: Benchmark
benchDListUtf8 = bench "utf8 + renderTableD maxiTable" $
  nf (L.length . B.toLazyByteString . B.stringUtf8 . D.toList . renderTableD) maxiTable


------------------------------------------------------------------------------
-- utf8-string and utf8-light
------------------------------------------------------------------------------

-- 4.12 ms
benchDListUtf8Light :: Benchmark
benchDListUtf8Light = bench "utf8-light + renderTable maxiTable" $
  whnf (Utf8Light.encode . D.toList . renderTableD) maxiTable

{- Couldn't get utf8-string to work :-(

benchDListUtf8String :: Benchmark
benchDListUtf8String = bench "utf8-light + renderTable maxiTable" $
  whnf (Utf8String.toRep . encode .
        D.toList . renderTableD) maxiTable
  where
    encode :: String -> Utf8String.UTF8 S.ByteString
    encode = Utf8String.fromString
-}

------------------------------------------------------------------------------
-- Data.Binary.Builder based rendering
------------------------------------------------------------------------------

-- Note that as of binary-0.6.0.0 the binary builder is the same as the one
-- provided by the bytestring library.
--
-- {-# INLINE char8BinB #-}
-- char8BinB :: Char -> BinB.Builder
-- char8BinB = BinB.singleton . fromIntegral . ord
--
-- renderStringBinB :: String -> BinB.Builder
-- renderStringBinB cs = char8BinB '"' <> foldMap escape cs <> char8BinB '"'
--   where
--     escape '\\' = char8BinB '\\' <> char8BinB '\\'
--     escape '\"' = char8BinB '\\' <> char8BinB '"'
--     escape c    = char8BinB c
--
-- renderCellBinB :: Cell -> BinB.Builder
-- renderCellBinB (StringC cs) = renderStringBinB cs
-- renderCellBinB (IntC i)     = B.intDec i
--
-- renderRowBinB :: Row -> BinB.Builder
-- renderRowBinB []     = mempty
-- renderRowBinB (c:cs) =
--     renderCellBinB c <> mconcat [ char8BinB ',' <> renderCellBinB c' | c' <- cs ]
--
-- renderTableBinB :: Table -> BinB.Builder
-- renderTableBinB rs = mconcat [renderRowBinB r <> char8BinB '\n' | r <- rs]
--
-- -- 1.22 ms
-- benchBinaryBuilderChar8 :: Benchmark
-- benchBinaryBuilderChar8 = bench "char8 + renderTableBinB maxiTable" $
--   nf (L.length . BinB.toLazyByteString . renderTableBinB) maxiTable
--

------------------------------------------------------------------------------
-- Text Builder
------------------------------------------------------------------------------

renderStringTB :: String -> TB.Builder
renderStringTB cs = TB.singleton '"' <> foldMap escape cs <> TB.singleton '"'
  where
    escape '\\' = "\\\\"
    escape '\"' = "\\\""
    escape c    = TB.singleton c

renderCellTB :: Cell -> TB.Builder
renderCellTB (StringC cs) = renderStringTB cs
renderCellTB (IntC i)     = TB.decimal i

renderRowTB :: Row -> TB.Builder
renderRowTB []     = mempty
renderRowTB (c:cs) =
    renderCellTB c <> mconcat [ TB.singleton ',' <> renderCellTB c' | c' <- cs ]

renderTableTB :: Table -> TB.Builder
renderTableTB rs = mconcat [renderRowTB r <> TB.singleton '\n' | r <- rs]

-- 0.95 ms
benchTextBuilder :: Benchmark
benchTextBuilder = bench "renderTableTB maxiTable" $
  nf (TL.length . TB.toLazyText . renderTableTB) maxiTable

-- 1.10 ms
benchTextBuilderUtf8 :: Benchmark
benchTextBuilderUtf8 = bench "utf8 + renderTableTB maxiTable" $
  nf (L.length . TL.encodeUtf8 . TB.toLazyText . renderTableTB) maxiTable



{- On a Core 2 Duo 2.2 GHz running a 32-bit Linux:


touching all data:                 0.25 ms
string rendering:                  1.36 ms
string rendering + utf8 encoding:  1.36 ms
DList rendering  + utf8 encoding:  0.91 ms
builder rendering (incl. utf8):    0.82 ms
builder + faster escaping:         0.65 ms

text builder:                      0.95 ms
text builder + utf8 encoding:      1.10 ms
binary builder + char8 (!!):       1.22 ms
DList render + utf8-light:         4.12 ms

How to improve further?
  * Use packed formats for string literals
    - fast memcpy  (that's what blaze-html does for tags)
    - using Text literals should also help


results from criterion:

benchmarking nf maxiTable
mean: 257.2927 us, lb 255.9210 us, ub 259.6692 us, ci 0.950
std dev: 9.026280 us, lb 5.887942 us, ub 12.76582 us, ci 0.950

benchmarking renderTable maxiTable
mean: 1.358458 ms, lb 1.356732 ms, ub 1.362377 ms, ci 0.950
std dev: 12.66932 us, lb 7.110377 us, ub 24.97397 us, ci 0.950

benchmarking utf8 + renderTable maxiTable
mean: 1.364343 ms, lb 1.362391 ms, ub 1.366973 ms, ci 0.950
std dev: 11.65388 us, lb 9.094074 us, ub 17.47765 us, ci 0.950

benchmarking utf8 + renderTableD maxiTable
mean: 909.5255 us, lb 908.0049 us, ub 911.7639 us, ci 0.950
std dev: 9.434182 us, lb 6.906120 us, ub 15.43223 us, ci 0.950

benchmarking utf8-light + renderTable maxiTable
mean: 4.128315 ms, lb 4.121109 ms, ub 4.138436 ms, ci 0.950
std dev: 42.93755 us, lb 32.58115 us, ub 58.61780 us, ci 0.950

benchmarking char8 + renderTableBinB maxiTable
mean: 1.224156 ms, lb 1.222510 ms, ub 1.226101 ms, ci 0.950
std dev: 9.046150 us, lb 7.568433 us, ub 11.74996 us, ci 0.950

benchmarking renderTableTB maxiTable
mean: 954.8066 us, lb 953.6650 us, ub 957.0134 us, ci 0.950
std dev: 7.763098 us, lb 5.072194 us, ub 14.09216 us, ci 0.950

benchmarking utf8 + renderTableTB maxiTable
mean: 1.095913 ms, lb 1.094811 ms, ub 1.098280 ms, ci 0.950
std dev: 7.865781 us, lb 4.189907 us, ub 15.24606 us, ci 0.950

benchmarking utf8 + renderTableB maxiTable
mean: 818.0223 us, lb 816.5118 us, ub 819.9397 us, ci 0.950
std dev: 8.603917 us, lb 6.764347 us, ub 12.29236 us, ci 0.950

benchmarking utf8 + renderTableBE maxiTable
mean: 646.5248 us, lb 645.3735 us, ub 648.2405 us, ci 0.950
std dev: 7.147889 us, lb 5.222494 us, ub 11.82482 us, ci 0.950

-}



{- Conclusion:
 -------------

 * Whenever generating a sequence of bytes: use the 'Builder' type

   => chunks can always be kept large; impossible when exporting only
      a strict/lazy bytestring interface.

   => filtering/mapping lazy bytestrings now automatically defragments
      the output and guarantees a large chunk size.


 * Status of work: API complete, documentation needs more reviewing.


 * Bounded encodings: safely exploiting low-level optimizations

   => a performance advantage on other outputstream-libraries?


                           ---------------
                           - Questions ? -
                           ---------------

-}




{- Implementation outline:
 ------------------------

data BufferRange = BufferRange {-# UNPACK #-} !(Ptr Word8)  -- First byte of range
                               {-# UNPACK #-} !(Ptr Word8)  -- First byte /after/ range

newtype BuildStep a =
    BuildStep { runBuildStep :: BufferRange -> IO (BuildSignal a) }

data BuildSignal a =
    Done             !(Ptr Word8)    -- next free byte in current buffer
                     a               -- return value
  | BufferFull
                     !Int            -- minimal size of next buffer
                     !(Ptr Word8)    -- next free byte in current buffer
                     !(BuildStep a)  -- continuation to call on next buffer
  | InsertByteString
                     !(Ptr Word8)    -- next free byte in current buffer
                     !S.ByteString   -- bytestring to insert directly
                     !(BuildStep a)  -- continuation to call on next buffer


-- | A "difference list" of build-steps.
newtype Builder = Builder (forall r. BuildStep r -> BuildStep r)


-- | The corresponding "Writer" monad.
newtype Put a = Put { unPut :: forall r. (a -> BuildStep r) -> BuildStep r }


-}

results.txt

$ ./benchcsv
Encoding the maxiTable
Total length in bytes: 112500
Chunk lengths: [4077,32745,32734,32749,10195]

benchmarking nf maxiTable
time                 829.4 μs   (820.4 μs .. 839.1 μs)
                     0.999 R²   (0.998 R² .. 0.999 R²)
mean                 831.2 μs   (825.0 μs .. 837.6 μs)
std dev              21.07 μs   (18.25 μs .. 24.76 μs)
variance introduced by outliers: 15% (moderately inflated)

benchmarking utf8 + renderTableB maxiTable + lazy
time                 1.368 ms   (1.344 ms .. 1.394 ms)
                     0.998 R²   (0.996 R² .. 0.999 R²)
mean                 1.380 ms   (1.364 ms .. 1.397 ms)
std dev              52.03 μs   (43.18 μs .. 69.63 μs)
variance introduced by outliers: 25% (moderately inflated)

benchmarking utf8 + renderTableB maxiTable  + streaming 
time                 1.335 ms   (1.320 ms .. 1.352 ms)
                     0.999 R²   (0.998 R² .. 0.999 R²)
mean                 1.341 ms   (1.327 ms .. 1.358 ms)
std dev              51.37 μs   (41.60 μs .. 67.58 μs)
variance introduced by outliers: 26% (moderately inflated)

benchmarking utf8 + renderTableBE maxiTable + lazy
time                 1.350 ms   (1.330 ms .. 1.364 ms)
                     0.998 R²   (0.997 R² .. 0.999 R²)
mean                 1.373 ms   (1.358 ms .. 1.390 ms)
std dev              53.87 μs   (45.07 μs .. 64.71 μs)
variance introduced by outliers: 26% (moderately inflated)

benchmarking utf8 + renderTableBE maxiTable  + streaming 
time                 1.388 ms   (1.369 ms .. 1.411 ms)
                     0.998 R²   (0.997 R² .. 0.999 R²)
mean                 1.404 ms   (1.388 ms .. 1.423 ms)
std dev              58.09 μs   (48.06 μs .. 72.49 μs)
variance introduced by outliers: 28% (moderately inflated)