store: a new and efficient binary serialization library

24 May 2016 Michael Sloan

A couple months ago, Michael Snoyman wrote a blogpost describing an experiment in an efficient implementation of binary serialization. Since then, we've developed this approach into a new package for efficient serialization of Haskell datatypes. I'm happy to announce that today we are putting out the initial release of our new new store package!

The store package takes a different approach than most prior serialization packages, in that performance is prioritized over other concerns. In particular, we do not make many guarantees about binary compatibility, and instead favor machine representations. For example, the binary and cereal packages use big endian encodings for numbers, whereas x86 machines use little endian. This means that to encode + decode numbers on an x86 machine, those packages end up swapping all of the individual bytes around twice!

To serialize a value, store first computes its size and allocates a properly sized ByteString. This keeps the serialization logic simple and fast, rather than mixing in logic to allocate new buffers. For datatypes that need to visit many values to compute their size, this can be inefficient - the datatype is traversed once to compute the size and once to do the serialization. However, for datatypes with constant size, or vectors of datatypes with constant size, it is possible to very quickly compute the total required size. List / set / map-like Store instances all implement this optimization when their elements have constant size.

store comes with instances for most datatypes from base, vector, bytestring, text, containers, and time. You can also use either GHC generics or Template Haskell to derive efficient instances for your datatypes.

Benchmark Results

I updated the serial-bench with store. Happily, store is even faster than any of the implementations we had in the benchmark.

serial-bench results

See the detailed report here. Note that the x-axis is measured in micro-seconds taken to serialize a 100 element Vector where each element occupies at least 17 bytes. store is actually performing this operations in the sub-microseconds (431ns to encode, 906ns to decode). The results for binary have been omitted from this graph as it blows out the x-axis scale, taking around 8 times longer than cereal, nearly 100x slower than store)

We could actually write a benchmark even more favorable to store, if we used storable or unboxed vectors! In that case, store essentially implements a memcpy.

Speeding up stack builds

Now, the benchmark is biased towards the usecase we are concerned with - serializing a Vector of a small datatype which always takes up the same amount of space. store was designed with this variety of usecase in mind, so naturally it excels in this benchmark. But lets say we choose a case that isn't exactly store's strongsuit, how well does it perform? In our experiments, it seems that store does a darn good job of that too!

The development version of stack now uses store for serializing caches of info needed by the build.

With store (~0.082 seconds):

2016-05-23 19:52:06.964518: [debug] Trying to decode /home/mgsloan/.stack/indices/Hackage/00-index.cache @(stack_I9M2eJwnG6d3686aQ2OkVk:Data.Store.VersionTagged src/Data/Store/VersionTagged.hs:49:5)
2016-05-23 19:52:07.046851: [debug] Success decoding /home/mgsloan/.stack/indices/Hackage/00-index.cache @(stack_I9M2eJwnG6d3686aQ2OkVk:Data.Store.VersionTagged src/Data/Store/VersionTagged.hs:58:13)

21210280 bytes

With binary (~0.197 seconds):

2016-05-23 20:22:29.855724: [debug] Trying to decode /home/mgsloan/.stack/indices/Hackage/00-index.cache @(stack_4Jm00qpelFc1pPl4KgrPav:Data.Binary.VersionTagged src/Data/Binary/VersionTagged.hs:55:5)
2016-05-23 20:22:30.053367: [debug] Success decoding /home/mgsloan/.stack/indices/Hackage/00-index.cache @(stack_4Jm00qpelFc1pPl4KgrPav:Data.Binary.VersionTagged src/Data/Binary/VersionTagged.hs:64:13)

20491950 bytes

So this part of stack is now twice as fast!


Beyond the core of store's functionality, this initial release also provides:

  • Data.Store.Streaming - functions for using Store for streaming serialization with conduit. This makes it so that you don't need to have everything in memory at once when serializing / deserializing. For applications involving lots of data, this can essential to having reasonable performance, or even functioning at all.

    This allows us to recoup the benefits of lazy serialization, without paying for the overhead when we don't need it. This approach is also more explicit / manual with regards to the laziness - the user must determine how their data will be streamed into chunks.

  • Data.Store.TypeHash, which provides utilities for computing hashes based on the structural definitions of datatypes. The purpose of this is to provide a mechanism for tagging serialized data in such a way that deserialization issues can be anticipated.

    This is included in the store package for a couple reasons:

    1. It is quite handy to include these hashes with your encoded datatypes. The assumption is that any structural differences are likely to correspond with serialization incompatibilities. This is particularly true when the generics / TH deriving is used rather than custom instances.

    2. It uses store on Template Haskell types in order to compute a ByteString. This allows us to directly use cryptographic hashes from the cryptohash package to get a hash of the type info.

  • Data.Store.TH not only provides a means to derive Store instances for your datatypes, but it also provides utilities for checking them via smallcheck and hspec. This makes it easy to check that all of your datatypes do indeed serialize properly.

These extras were the more recently added parts of store, and so are likely to change quite a bit from the current API. The entirety of store is quite new, and so is also subject to API change while it stabilizes. That said, we encourage you to give it a try for your application!

TH cleverness

Usually, we directly use Storable instances to implement Store. In functionality, Storable is very similar to Store. The key difference is that Store instances can take up a variable amount of size, whereas Storable types must use a constant number of bytes. The store package also provides the convenience of Peek and Poke monads, so defining custom Store instances is quite a bit more convenient

Data.Store.TH.Internal defines a function deriveManyStoreFromStorable, which does the following:

  • Reifies all Store instances
  • Reifies all Storable instances.
  • Implements Store instances for all Storable instances

In the future, store will likely provide such a function for users, which restricts it to only deriving Store instances for types in the current package or current module. For now, this is just internal convenience.

I noticed that the Storable instance for Bool is a bit wasteful with its bytes. Rather inexplicably, perhaps due to alignment concerns, it takes up a whopping 4 bytes to represent a single bit of info:

instance Storable Bool where
   sizeOf _          = sizeOf (undefined::HTYPE_INT)
   alignment _       = alignment (undefined::HTYPE_INT)
   peekElemOff p i   = liftM (/= (0::HTYPE_INT)) $ peekElemOff (castPtr p) i
   pokeElemOff p i x = pokeElemOff (castPtr p) i (if x then 1 else 0::HTYPE_INT)

We'd prefer to just use a single byte. Since deriveManyStoreFromStorable skips types that already have Store instances, all I needed to do was define our own instance for Bool. To do this, I used the derive function from the new th-utilities package (blogpost pending!), to define an instance for Bool:

$($(derive [d|
    instance Deriving (Store Bool)

This is a bit of a magical incantation - it runs code at compiletime which generates an efficient instance Store Bool where .... We could also use generic deriving, and rely on the method defaults to just write instance Store Bool. However, this can be less efficient, because the generics instances will yield a VarSize for its size, whereas the TH instance is smart enough to yield ConstSize. In practice, this is the difference between having an O(1) implementation for size :: Size (Vector MyADT), and having an O(n) implementation. The O(1) implementation just multiplies the element size by the length, whereas the O(n) implementation needs to ask each element for its size.

comments powered by Disqus

Copyright © 2013-2017 FP Complete Corp. All rights reserved