11 releases (7 breaking)
0.8.0 | Feb 24, 2020 |
---|---|
0.7.1 | Feb 24, 2020 |
0.6.0 | Feb 23, 2020 |
0.5.0 | Feb 22, 2020 |
0.1.1 | Feb 20, 2020 |
#1680 in Data structures
48KB
914 lines
Bucket Vector
Docs | Crates.io |
---|---|
⚠️ Caution
USE WITH CAUTION
As of now this crate has not yet been battle tested or benchmarked to an extend where the author would recommend general production usage. Please file bugs, suggestions or enhancements to the issue tracker of this repository.
Description
A vector-like data structure that organizes its elements into a set of buckets of fixed-capacity in order to guarantee that mutations to the bucket vector never moves elements and thus invalidates references to them.
This is comparable to a Vec<Box<T>>
but a lot more efficient.
Configs
The BucketVecConfig
trait allows to customize the internal structure of your
BucketVec
. This allows users to fine-tune their BucketVec
for particular
use cases.
The trait mainly controls the capacity of the first bucket and the growth rate of the capacity of new buckets.
The default DefaultConfig
tries to balance out the different interests
between start capacity and growth rate.
Under the Hood
The BucketVec
is really just a vector of Bucket
instances.
Whenever an element is pushed to the BucketVec
the element is pushed onto
the last Bucket
if it isn't filled, yet.
If the last Bucket
is filled a new Bucket
is pushed onto the BucketVec
with a new capacity determined by the used bucket vector configuration.
This way the BucketVec
never moves elements around upon inserting new elements
in order to preserve references. When a normal Vec
is modified it can potentially
invalidate references because of reallocation of the internal buffer which
might cause severe bugs if references to the internal elements are stored
outside the Vec
. Note that normally Rust prevents such situations so the
BucketVec
is mainly used in the area of unsafe
Rust where a developer
actively decides that they want or need pinned references into another data
structure.
For the same reasons as stated above the BucketVec
does not allow to remove
or swap elements.
Example
Looking at an example BucketVec<i32>
with the following configuration:
start_capacity := 1
growth_rate := 2
We have already pushed the elements A
,.., K
onto it.
[ [A], [B, C], [D, E, F, G], [H, I, J, K, _, _, _, _] ]
Where _
refers to a vacant bucket entry.
Pushing another L
,.., O
onto the same BucketVec
results in:
[ [A], [B, C], [D, E, F, G], [H, I, J, K, L, M, N, O] ]
So we are full on capacity for all buckets.
The next time we push another element onto the BucketVec
it will create a new Bucket
with a capacity of 16
since growth_rate == 2
and our current latest bucket already has a capacity of 8
.
[ [A], [B, C], [D, E, F, G], [H, I, J, K, L, M, N, O], [P, 15 x _] ]
Where 15 x _
denotes 15 consecutive vacant entries.
Benchmarks
BucketVec
fullfills its role in being a replacement for situations where
a Vec<Box<T>>
is the naive go-to solution.
The benchmark suite is still small and not super expressive but already provides
some insights in where BucketVec
already performs pretty well and where it
could improve.
Benchmarks have been run on a Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz.
Note that for every of the benchmark groups (push
, get
and iter
) the
most efficient configuration for the BucketVec
has been chosen.
Some benchmarks (get
) have shown significant difference between configs.
The following is the output of a benchmark run:
bucket_vec::push/10000 time: [43.647 us 43.861 us 44.108 us]
bucket_vec::push_get/10000 time: [48.872 us 49.396 us 49.834 us]
vec_box::push/10000 time: [405.37 us 410.91 us 417.20 us]
vec_value::push/10000 time: [25.826 us 25.915 us 26.020 us]
bucket_vec::get/10000 time: [17.369 us 17.416 us 17.470 us]
vec_box::get/10000 time: [14.446 us 14.485 us 14.537 us]
vec_value::get/10000 time: [8.7939 us 8.8105 us 8.8300 us]
bucket_vec::iter/10000 time: [4.4195 us 4.4316 us 4.4454 us]
vec_box::iter/10000 time: [9.5925 us 9.6246 us 9.6610 us]
vec_value::iter/10000 time: [3.5955 us 3.6043 us 3.6142 us]
bucket_vec::iter.rev()/10000 time: [3.9804 us 3.9957 us 4.0144 us]
vec_box::iter.rev()/10000 time: [9.9677 us 9.9980 us 10.033 us]
vec_value::iter.rev()/10000 time: [3.5827 us 3.5944 us 3.6080 us]
bucket_vec::iter_mut/10000 time: [5.0533 us 5.0710 us 5.0909 us]
vec_box::iter_mut/10000 time: [13.425 us 13.845 us 14.203 us]
vec_value::iter_mut/10000 time: [4.0172 us 4.0473 us 4.0820 us]
It can be seen that BucketVec
greatly outperforms Vec<Box<_>>
on
push
, iter
and iter().rev()
benchmarks.
Also BucketVec
is approximately 50% slower than the Vec<_>
which is the
theoretical optimum that unfortunately doesn't solve the underlying problem.
Authors & Credits
Author: Robin Freyler (github.com/Robbepop)
Special thanks to Niklas Tittjung (github.com/lugino-emeritus) who helped me a lot with some internal formulae.
License
Licensed under either of
- Apache license, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
- MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)
at your option.
Dual licence:
Dependencies
~380–730KB
~16K SLoC