mason: Fast and extensible bytestring builder

[ bsd3, data, library ] [ Propose Tags ]

See README.md


[Skip to Readme]

Downloads

Maintainer's Corner

Package maintainers

For package maintainers and hackage trustees

Candidates

Versions [RSS] 0, 0.1, 0.2, 0.2.1, 0.2.2, 0.2.3, 0.2.4, 0.2.5, 0.2.6
Change log CHANGELOG.md
Dependencies base (>=4.12.0.0 && <5), bytestring, ghc-prim, integer-gmp, network (>=2.7 && <3.2), text [details]
License BSD-3-Clause
Copyright 2019 Fumiaki Kinoshita, Don Stewart 2005-2009, Duncan Coutts 2006-2015, David Roundy 2003-2005, Jasper Van der Jeugt 2010, Simon Meier 2010-2013, Ben Gamari 2017
Author Fumiaki Kinoshita
Maintainer fumiexcel@gmail.com
Category Data
Bug tracker https://github.com/fumieval/mason/issues
Uploaded by FumiakiKinoshita at 2019-12-06T12:55:22Z
Distributions LTSHaskell:0.2.6, NixOS:0.2.6, Stackage:0.2.6
Reverse Dependencies 1 direct, 1 indirect [details]
Downloads 2316 total (33 in the last 30 days)
Rating (no votes yet) [estimated by Bayesian average]
Your Rating
  • λ
  • λ
  • λ
Status Docs uploaded by user
Build status unknown [no reports yet]

Readme for mason-0.1

[back to package description]

mason: alacritous builder library

Build Status Hackage

mason is a builder & IO library.

  • Fast: much faster than bytestring's Builder.
  • Extensible: Builders can be consumed in a user-defined way.
  • Hackable: Low-level APIs are exposed. It's easy to plug in even pointer-level operations.

Mason.Builder has API mostly compatible with Data.ByteString.Builder but there are some additions to the original API:

  • toStrictByteString produces a strict ByteString directly.
  • hPutBuilderLen writes a builder to a handle and returns the number of bytes.
  • sendBuilder sends the content of Builder over a socket.

Usage

Replace Data.ByteString.Builder with Mason.Builder. Note that if you have Builder in the type signature, you'll need RankNTypes extensions because of the design explained below.

Performance

As long as the code is optimised, mason's builder can be very fast (twice or more as bytestring). Make sure that functions returning Builders are well inlined.

Serialisation of JSON-like structure:

mason/hPutBuilder                        mean 274.7 μs  ( +- 49.40 μs  )
fast-builder/hPutBuilder                 mean 399.9 μs  ( +- 76.05 μs  )
bytestring/hPutBuilder                   mean 335.1 μs  ( +- 86.96 μs  )
mason/toStrictByteString                 mean 106.6 μs  ( +- 6.680 μs  )
fast-builder/toStrictByteString          mean 254.8 μs  ( +- 31.64 μs  )
bytestring/toLazyByteString              mean 283.3 μs  ( +- 24.26 μs  )
mason/toLazyByteString                   mean 127.2 μs  ( +- 25.86 μs  )
fast-builder/toLazyByteString            mean 249.0 μs  ( +- 25.60 μs  )
bytestring/toLazyByteString              mean 263.4 μs  ( +- 9.401 μs  )

In the same benchmark application, the allocation footprint of mason is feathery.

toStrictByteString
mason           291,112    0
fast-builder    991,016    0
bytestring    1,158,584    0 (toStrict . toLazyByteString)

toLazyByteString
Case          Allocated  GCs
mason           228,936    0
fast-builder    903,752    0
bytestring    1,101,448    0

doubleDec employs Grisu3 which grants ~20x speedup over show-based implementation.

mason/double                             mean 116.2 ns  ( +- 6.654 ns  )
fast-builder/double                      mean 2.183 μs  ( +- 85.80 ns  )
bytestring/double                        mean 2.312 μs  ( +- 118.8 ns  )

TBD: more benchmarks

Architecture

Mason's builder is a function that takes a purpose-dependent environment and a buffer. There is little intermediate structure involved; almost everything runs in one pass. This design is inspired by fast-builder.

type Builder = forall s. Buildable s => BuilderFor s

newtype BuilderFor s = Builder { unBuilder :: s -> Buffer -> IO Buffer }

data Buffer = Buffer
  { bEnd :: {-# UNPACK #-} !(Ptr Word8) -- ^ end of the buffer (next to the last byte)
  , bCur :: {-# UNPACK #-} !(Ptr Word8) -- ^ current position
  }

class Buildable s where
  byteString :: B.ByteString -> BuilderFor s
  flush :: BuilderFor s
  allocate :: Int -> BuilderFor s

Instances of the Buildable class implement purpose-specific behaviour (e.g. exponentially allocate a buffer, flush to disk). This generic interface also allows creative uses of Builders such as on-the-fly compression.

Builder has a smart constructor called ensure:

ensure :: Int -> (Buffer -> IO Buffer) -> Builder

ensure n f secures at least n bytes in the buffer and passes the pointer to f. This gives rise to monoid homorphism; namely, ensure m f <> ensure n g will fuse into ensure (m + n) (f >=> g) so don't worry about the overhead of bound checking.

Creating your own primitives

The easiest way to create a new primitive is withPtr, a simplified version of ensure. This is quite convenient for calling foreign functions or anything low-level.

-- | Construct a 'Builder' from a "poke" function.
withPtr :: Int -- ^ number of bytes to allocate (if needed)
  -> (Ptr Word8 -> IO (Ptr Word8)) -- ^ return a next pointer after writing
  -> Builder

grisu v = withPtr 24 $ \ptr -> do
  n <- dtoa_grisu3 v ptr
  return $ plusPtr ptr (fromIntegral n)

foreign import ccall unsafe "static dtoa_grisu3"
  dtoa_grisu3 :: Double -> Ptr Word8 -> IO CInt