Crates.io | sdset |
lib.rs | sdset |
version | 0.4.0 |
source | src |
created_at | 2018-08-10 13:10:15.568235 |
updated_at | 2020-02-18 12:37:27.454277 |
description | Set operations for sorted and deduplicated slices. Much performances! Such Wow! |
homepage | |
repository | https://github.com/Kerollmops/sdset |
max_upload_size | |
id | 78684 |
size | 168,638 |
Set theory applied on sorted and deduplicated slices. Much performance! Such Wow!
API Documentation can be found on docs.rs.
sdset
stands for sorted-deduplicated-slices-set
which is a little bit too long.
Note about the tests, which are done on ranges of integer, if it ends with:
two_slices_big
, the first slice contains 0..100
and the second has 1..101
two_slices_big2
, the first contains 0..100
and the second has 51..151
two_slices_big3
, the first contains 0..100
and the second has 100..200
three_slices_big
, the first contains 0..100
, the second has 1..101
and the third has 2..102
three_slices_big2
, the first contains 0..100
, the second has 34..134
and the third has 67..167
three_slices_big3
, the first contains 0..100
, the second has 100..200
and the third has 200..300
These slices of runs of integer are useful when they overlap, we can see how performances changes when different parts of the slices overlaps.
For more informations on "Why is there no really big slices benchmarks ?", you can see my response on /r/rust.
To run the benchmarks you must enable the unstable
feature.
$ cargo bench --features unstable
Note that the sdset
set operations does not need many allocations so it starts with a serious advantage. For more information you can see the benchmarks variance.
_btree
are benchmarks that uses two or three BTreeSet
s which contains runs of integers (see above), the BTreeSet
s creations are not taken into account. The set operations are done on these sets and the result is accumulated in a final Vec
.
_fnv
are benchmarks that uses two or three HashSet
s which contains runs of integers (see above), it uses a custom Hasher
named fnv
that is specialized for little values like integers, the HashSet
s creations are not taken into account. The set operations are done on these sets and the result is accumulated in a final Vec
.
The _vec
benchmarks are available for the union set operation only, it consist of a Vec
which is populated with the elements of two or three slices (see above), sorted and deduplicated.
The duo
and multi
measurements are the implementations that are part of this crate, the first one can only do set operations on two sets and the second one can be used for any given number of sets.
Histograms can be generated using the benchmarks by executing the following command:
$ export CARGO_BENCH_CMD='cargo bench --features unstable'
$ ./gen_graphs.sh xxx.bench
This is much more easier to read statistics and to see how sdset
is more performant on already sorted and deduplicated slices than any other kind of collection.