An implementation of HAMT(Hash Array Mapped Trie, Bagwell) in Swift. Eonil, May 2019.
Use HAMT
type. This type provides these features.
- Hash-based key-value storage.
- All of read/write/copy
amortized O(log(n))
time up to certain number of elements (see Performance section), andO(n)
at worst. - Full "Copy-on-Write" behavior with minimal amount of copying.
The type provides these dictionary-like interfaces.
- Conformance to
Sequence
protocol. - Conformance to
Equatable
protocol. isEmpty: Bool
count: Int
subscript(Key) -> Value? { get set }
subscript(Key, default: @autoclosure () -> Value) -> Value { get set }
keys: Sequence
values: Sequence
These features are not supported (maybe yet).
Index
and index based look-up and iteration.- Any other collection protocol conformance.
As like most Swift datastrctures, HAMT
is also fully CoW compliant. This means
each copy of same HAMT tree shares data as much as much possible. Regardles
of how many copies you make, HAMT shares most portion of tree with all other
copies.
HAMT
type in this library is designed to be used as
persistent datastructure.
In copy-persistent scenario under 64-bit environment,
HAMT
provides near constant time (O(log(n)) & max depth=10, -> O(10)
) single element read/write/copy
performance up to hash resolution limit ((2^6)^10
items) regardless of contained item
count if hash function is well distributed. Also new copy does not take extra space unless
gets mutated. Copy with single element mutation takes O(log(n))
extra time and space.
On the other hand, copying Swift.Dictionary
takes O(n)
time and extra space.
Instead, single element read/write of HAMT
is about 2x/50x times slower
than ephemeral Swift.Dictionary
for random 64-bit integer keys and values.
Note that "operation count" in above graph is accumulated number.
Here's another performance comparison with copying B-Tree.
Naive Swift.Dictionary
is not drawn here because read/write performance
is same with ephemeral one, and copying it takes too much time and didn't finish.
For small dataset, naive copying of Swift.Dictionary
works better, but as
copying cost increases linearly, it is no longer efficient after 1,000 items.
Therefore, HAMT
is better if you need a hash-based persistent associative array
data structure that can grow more than several thousands.
B-Tree shows better write performance overall. HAMT performs better after 100K
items, but it doesn't seem to be really practical numbers. And it requires keys
to be Comparable
. By the way, as B-Tree doesn't have hash collision, it'll show
more reliable performance.
HAMT
type is implemented using PD5Bucket64
internally.
PD5Bucket64
type provides all additional properties for testing and
validation.
PD4
type was an implementation of hash-trie, and deprecated due to
high rate of wasted memory. PD5
implements HAMT and shows nearly
same performance with PD4
with far less memory consumption.
I used PD5
prefix for convenience only for internals. Public major type
name is HAMT
, and internal types all use PD5
prefixed. If I implement
a next version of algorithm, it'll be named as PD6
.
If once implementation gets stabilized, maybe I'll rename all PDx
prefixes
to HAMT
someday.
At this point, only 64-bit platforms are considered. It should work on 32-bit platforms but less performance and have not been tested.
If you link this library, you'll notice the performance is not good as shown in the graph. As like Károly Lőrentey clarified, it's because Swift compiler does not inline and optimize over externally linked functions. You can compile HAMT source code with your code together in same module to archive best possible performance.
-
See also "B-Tree for Swift" by Károly Lőrentey if you need sorted associative array.
-
For more information about HAMT, see the paper by Phil Bagwell.
Sending contribution means implicit agreement to redistribute your contribution under "MIT License".
This code is licensed under "MIT License". Copyright Eonil, Hoon H.. 2019. All rights reserved.