FP Complete


As part of our high-performance computing work, I recently found myself in need of some fast mutable containers. The code is now available on both Hackage and Stackage. The code is pretty young, and is open to design changes still. That said, the currently released version (0.2.0) is well tested and performs fairly well. If there are ideas for improvement, please let me know!

Below is the content of the README file, which gives a good overview of the package, as well as benchmark numbers and test coverage statistics (spoiler: 100%). As always, you can see the README on Github or on Stackage.


One of Haskell’s strengths is immutable data structures. These structures make it easier to reason about code, simplify concurrency and parallelism, and in some cases can improve performance by allowing sharing. However, there are still classes of problems where mutable data structures can both be more convenient, and provide a performance boost. This library is meant to provide such structures in a performant, well tested way. It also provides a simple abstraction over such data structures via typeclasses.

Before anything else, let me provide the caveats of this package:

We’ll first talk about the general approach to APIs in this package. Next, there are two main sets of abstractions provided, which we’ll cover in the following two sections, along with their concrete implementations. Finally, we’ll cover benchmarks.

API structure

The API takes heavy advantage of the PrimMonad typeclass from the primitive package. This allows our data structures to work in both IO and ST code. Each data structure has an associated type, MCState, which gives the primitive state for that structure. For example, in the case of IORef, that state is RealWorld, whereas for STRef s, it would be s. This associated type is quite similar to the PrimState associated type from primitive, and in many type signatures you’ll see an equality constraint along the lines of:

PrimState m ~ MCState c

For those who are wondering, MCState stands for “mutable container state.”

All actions are part of a typeclass, which allows for generic access to different types of structures quite easily. In addition, we provide type hint functions, such as asIORef, which can help specify types when using such generic functions. For example, a common idiom might be:

ioref <- fmap asIORef $ newRef someVal

Wherever possible, we stick to well accepted naming and type signature standards. For example, note how closely modifyRef and modifyRef' match modifyIORef and modifyIORef'.

Single cell references

The base package provides both IORef and STRef as boxed mutable references, for storing a single value. The primitive package also provides MutVar, which generalizes over both of those and works for any PrimMonad instance. The MutableRef typeclass in this package abstracts over all three of those. It has two associated types: MCState for the primitive state, and RefElement to specify what is contained by the reference.

You may be wondering: why not just take the reference as a type parameter? That wouldn’t allow us to have monomorphic reference types, which may be useful under some circumstances. This is a similar motivation to how the mono-traversable package works.

In addition to providing an abstraction over IORef, STRef, and MutVar, this package provides four additional single-cell mutable references. URef, SRef, and BRef all contain a 1-length mutable vector under the surface, which is unboxed, storable, and boxed, respectively. The advantage of the first two over boxed standard boxed references is that it can avoid a significant amount of allocation overhead. See the relevant Stack Overflow discussion and the benchmarks below.

While BRef doesn’t give this same advantage (since the values are still boxed), it was trivial to include it along with the other two, and does actually demonstrate a performance advantage. Unlike URef and SRef, there is no restriction on the type of value it can store.

The final reference type is PRef. Unlike the other three mentioned, it doesn’t use vectors at all, but instead drops down directly to a mutable bytearray to store values. This means it has slightly less overhead (no need to store the size of the vector), but also restricts the types of things that can be stored (only instances of Prim).

You should benchmark your program to determine the most efficient reference type, but generally speaking PRef will be most performant, followed by URef and SRef, and finally BRef.

Collections

Collections allow you to push and pop values to the beginning and end of themselves. Since different data structures allow different operations, each operation goes into its own typeclass, appropriately named MutablePushFront, MutablePushBack, MutablePopFront, and MutablePopBack. There is also a parent typeclass MutableCollection which provides:

  1. The CollElement associated type to indicate what kinds of values are in the collection.
  2. The newColl function to create a new, empty collection.

The mono-traversable package provides a typeclass IsSequence which abstracts over sequence-like things. In particular, it provides operations for cons, snoc, uncons, and unsnoc. Using this abstraction, we can provide an instance for all of the typeclasses listed above for any mutable reference containing an instance of IsSequence, e.g. IORef [Int] or BRef s (Seq Double).

Note that the performance of some of these combinations is terrible. In particular, pushBack or popBack on a list requires traversing the entire list, and any push operations on a Vector requires copying the entire contents of the vector. Caveat emptor! If you must use one of these structures, it’s highly recommended to use Seq, which gives the best overall performance.

However, in addition to these instances, this package also provides two additional data structures: double-ended queues and doubly-linked lists. The former is based around mutable vectors, and therefore as unboxed (UDeque), storable (SDeque), and boxed (BDeque) variants. Doubly-linked lists have no such variety, and are simply DLists.

For general purpose queue-like structures, UDeque or SDeque is likely to give you best performance. As usual, benchmark your own program to be certain, and see the benchmark results below.

Benchmark results

The following benchmarks were performed on January 7, 2015, against version 0.2.0.

Ref benchmark

benchmarking IORef
time                 4.322 μs   (4.322 μs .. 4.323 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 4.322 μs   (4.322 μs .. 4.323 μs)
std dev              1.401 ns   (1.114 ns .. 1.802 ns)

benchmarking STRef
time                 4.484 μs   (4.484 μs .. 4.485 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 4.484 μs   (4.484 μs .. 4.484 μs)
std dev              941.0 ps   (748.5 ps .. 1.164 ns)

benchmarking MutVar
time                 4.482 μs   (4.482 μs .. 4.483 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 4.482 μs   (4.482 μs .. 4.483 μs)
std dev              843.2 ps   (707.9 ps .. 1.003 ns)

benchmarking URef
time                 2.020 μs   (2.019 μs .. 2.020 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 2.020 μs   (2.019 μs .. 2.020 μs)
std dev              955.2 ps   (592.2 ps .. 1.421 ns)

benchmarking PRef
time                 2.015 μs   (2.014 μs .. 2.015 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 2.014 μs   (2.014 μs .. 2.015 μs)
std dev              901.3 ps   (562.8 ps .. 1.238 ns)

benchmarking SRef
time                 2.231 μs   (2.230 μs .. 2.232 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 2.231 μs   (2.230 μs .. 2.231 μs)
std dev              1.938 ns   (1.589 ns .. 2.395 ns)

benchmarking BRef
time                 4.279 μs   (4.279 μs .. 4.279 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 4.279 μs   (4.279 μs .. 4.279 μs)
std dev              1.281 ns   (1.016 ns .. 1.653 ns)

Deque benchmark

time                 8.371 ms   (8.362 ms .. 8.382 ms)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 8.386 ms   (8.378 ms .. 8.398 ms)
std dev              29.25 μs   (20.73 μs .. 42.47 μs)

benchmarking IORef (Seq Int)
time                 142.9 μs   (142.7 μs .. 143.1 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 142.7 μs   (142.6 μs .. 142.9 μs)
std dev              542.8 ns   (426.5 ns .. 697.0 ns)

benchmarking UDeque
time                 107.5 μs   (107.4 μs .. 107.6 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 107.5 μs   (107.4 μs .. 107.6 μs)
std dev              227.4 ns   (171.8 ns .. 297.8 ns)

benchmarking SDeque
time                 97.82 μs   (97.76 μs .. 97.89 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 97.82 μs   (97.78 μs .. 97.89 μs)
std dev              169.5 ns   (110.6 ns .. 274.5 ns)

benchmarking BDeque
time                 113.5 μs   (113.4 μs .. 113.6 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 113.6 μs   (113.5 μs .. 113.7 μs)
std dev              300.4 ns   (221.8 ns .. 424.1 ns)

benchmarking DList
time                 156.5 μs   (156.3 μs .. 156.6 μs)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 156.4 μs   (156.3 μs .. 156.6 μs)
std dev              389.5 ns   (318.3 ns .. 502.8 ns)

Test coverage

As of version 0.2.0, this package has 100% test coverage. If you look at the report yourself, you’ll see some uncovered code; it’s just the automatically derived Show instance needed for QuickCheck inside the test suite itself.

Subscribe to our blog via email

Email subscriptions come from our Atom feed and are handled by Blogtrottr. You will only receive notifications of blog posts, and can unsubscribe any time.

Tagged