Commit graph

19 commits

Author SHA1 Message Date
266da7c69a [#8] hrw: Do not allocate for swap()/less() helpers
```
goos: linux
goarch: amd64
pkg: git.frostfs.info/TrueCloudLab/hrw
cpu: 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz
                                         │      4      │                  5                  │
                                         │   sec/op    │   sec/op     vs base                │
SortHashersByValue_Typed_fnv_10-8          309.2n ± 2%   294.4n ± 1%   -4.75% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8         2.306µ ± 1%   2.549µ ± 1%  +10.54% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8        21.73µ ± 1%   24.80µ ± 3%  +14.14% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8     347.1n ± 1%   334.8n ± 2%   -3.56% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8    2.668µ ± 1%   2.954µ ± 3%  +10.72% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8   2.673µ ± 1%   2.957µ ± 4%  +10.63% (p=0.000 n=10)
geomean                                    1.836µ        1.947µ        +6.01%

                                         │      4       │                  5                   │
                                         │     B/op     │     B/op      vs base                │
SortHashersByValue_Typed_fnv_10-8            216.0 ± 0%     144.0 ± 0%  -33.33% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8          1032.0 ± 0%     960.0 ± 0%   -6.98% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8        8.133Ki ± 0%   8.062Ki ± 0%   -0.86% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8       216.0 ± 0%     144.0 ± 0%  -33.33% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8     1032.0 ± 0%     960.0 ± 0%   -6.98% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8    1032.0 ± 0%     960.0 ± 0%   -6.98% (p=0.000 n=10)
geomean                                      867.8          730.1       -15.87%

                                         │     4      │                 5                  │
                                         │ allocs/op  │ allocs/op   vs base                │
SortHashersByValue_Typed_fnv_10-8          4.000 ± 0%   2.000 ± 0%  -50.00% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8         4.000 ± 0%   2.000 ± 0%  -50.00% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8        4.000 ± 0%   2.000 ± 0%  -50.00% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8     4.000 ± 0%   2.000 ± 0%  -50.00% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8    4.000 ± 0%   2.000 ± 0%  -50.00% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8   4.000 ± 0%   2.000 ± 0%  -50.00% (p=0.000 n=10)
geomean                                    4.000        2.000       -50.00%
```

Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
2023-06-02 10:56:23 +03:00
c52f74d8e1 [#8] hrw: Do not allocate 2 slices for sort
Currently we allocate `rule` and then create `dist` which depends on it.
In this commit we create `dist` directly.

```
goos: linux
goarch: amd64
pkg: git.frostfs.info/TrueCloudLab/hrw
cpu: 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz
                                         │      3      │                 4                  │
                                         │   sec/op    │   sec/op     vs base               │
SortHashersByValue_Typed_fnv_10-8          336.3n ± 3%   309.2n ± 2%  -8.06% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8         2.424µ ± 3%   2.306µ ± 1%  -4.87% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8        22.35µ ± 1%   21.73µ ± 1%  -2.75% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8     346.6n ± 3%   347.1n ± 1%       ~ (p=0.631 n=10)
SortHashersByWeightValueTyped_fnv_100-8    2.637µ ± 6%   2.668µ ± 1%       ~ (p=0.481 n=10)
SortHashersByWeightValueTyped_fnv_1000-8   2.609µ ± 4%   2.673µ ± 1%  +2.43% (p=0.000 n=10)
geomean                                    1.875µ        1.836µ       -2.06%

                                         │       3       │                  4                   │
                                         │     B/op      │     B/op      vs base                │
SortHashersByValue_Typed_fnv_10-8             296.0 ± 0%     216.0 ± 0%  -27.03% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8          1.883Ki ± 0%   1.008Ki ± 0%  -46.47% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8        16.133Ki ± 0%   8.133Ki ± 0%  -49.59% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8        296.0 ± 0%     216.0 ± 0%  -27.03% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8     1.883Ki ± 0%   1.008Ki ± 0%  -46.47% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8    1.883Ki ± 0%   1.008Ki ± 0%  -46.47% (p=0.000 n=10)
geomean                                     1.442Ki          867.8       -41.24%

                                         │     3      │                 4                  │
                                         │ allocs/op  │ allocs/op   vs base                │
SortHashersByValue_Typed_fnv_10-8          5.000 ± 0%   4.000 ± 0%  -20.00% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8         5.000 ± 0%   4.000 ± 0%  -20.00% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8        5.000 ± 0%   4.000 ± 0%  -20.00% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8     5.000 ± 0%   4.000 ± 0%  -20.00% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8    5.000 ± 0%   4.000 ± 0%  -20.00% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8   5.000 ± 0%   4.000 ± 0%  -20.00% (p=0.000 n=10)
geomean                                    5.000        4.000       -20.00%
```

Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
2023-06-02 10:53:44 +03:00
895ecf150f [#8] hrw: Inline swap() when slice is known
```
goos: linux
goarch: amd64
pkg: git.frostfs.info/TrueCloudLab/hrw
cpu: 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz
                                         │      2      │                 3                  │
                                         │   sec/op    │   sec/op     vs base               │
SortHashersByValue_Typed_fnv_10-8          368.5n ± 2%   336.3n ± 3%  -8.75% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8         2.411µ ± 4%   2.424µ ± 3%       ~ (p=0.853 n=10)
SortHashersByValue_Typed_fnv_1000-8        22.19µ ± 2%   22.35µ ± 1%       ~ (p=0.247 n=10)
SortHashersByWeightValueTyped_fnv_10-8     364.3n ± 2%   346.6n ± 3%  -4.86% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8    2.541µ ± 3%   2.637µ ± 6%       ~ (p=0.055 n=10)
SortHashersByWeightValueTyped_fnv_1000-8   2.483µ ± 1%   2.609µ ± 4%  +5.07% (p=0.003 n=10)
geomean                                    1.888µ        1.875µ       -0.71%

                                         │      2       │                  3                  │
                                         │     B/op     │     B/op      vs base               │
SortHashersByValue_Typed_fnv_10-8            312.0 ± 0%     296.0 ± 0%  -5.13% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8         1.898Ki ± 0%   1.883Ki ± 0%  -0.82% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8        16.15Ki ± 0%   16.13Ki ± 0%  -0.10% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8       312.0 ± 0%     296.0 ± 0%  -5.13% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8    1.898Ki ± 0%   1.883Ki ± 0%  -0.82% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8   1.898Ki ± 0%   1.883Ki ± 0%  -0.82% (p=0.000 n=10)
geomean                                    1.474Ki        1.442Ki       -2.16%

                                         │     2      │                 3                  │
                                         │ allocs/op  │ allocs/op   vs base                │
SortHashersByValue_Typed_fnv_10-8          6.000 ± 0%   5.000 ± 0%  -16.67% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8         6.000 ± 0%   5.000 ± 0%  -16.67% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8        6.000 ± 0%   5.000 ± 0%  -16.67% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8     6.000 ± 0%   5.000 ± 0%  -16.67% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8    6.000 ± 0%   5.000 ± 0%  -16.67% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8   6.000 ± 0%   5.000 ± 0%  -16.67% (p=0.000 n=10)
geomean                                    6.000        5.000       -16.67%
```

Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
2023-06-01 21:41:16 +03:00
213c105ac1 [#8] go.mod: Use faster murmur3 lib
Specifically, this line became possible, because of noescape annotations
for assembly.
```
./hrw.go:307:14: make([]byte, 8) does not escape
```

```
goos: linux
goarch: amd64
pkg: git.frostfs.info/TrueCloudLab/hrw
cpu: 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz
                                              │      1       │                   2                   │
                                              │    sec/op    │   sec/op     vs base                  │
SortHashersByValue_Typed_fnv_10-8               580.1n ±  1%   368.5n ± 2%  -36.47% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8              4.215µ ±  2%   2.411µ ± 4%  -42.79% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8             39.40µ ±  1%   22.19µ ± 2%  -43.68% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8          599.6n ±  2%   364.3n ± 2%  -39.25% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8         4.337µ ±  5%   2.541µ ± 3%  -41.41% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8        4.344µ ±  3%   2.483µ ± 1%  -42.84% (p=0.000 n=10)
geomean                                         4.400µ         1.888µ       -41.13%                ¹
¹ benchmark set differs from baseline; geomeans may not be comparable

                                              │      1       │                   2                    │
                                              │     B/op     │     B/op      vs base                  │
SortHashersByValue_Typed_fnv_10-8                 472.0 ± 0%     312.0 ± 0%  -33.90% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8              3.461Ki ± 0%   1.898Ki ± 0%  -45.15% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8             31.77Ki ± 0%   16.15Ki ± 0%  -49.18% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8            472.0 ± 0%     312.0 ± 0%  -33.90% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8         3.461Ki ± 0%   1.898Ki ± 0%  -45.15% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8        3.461Ki ± 0%   1.898Ki ± 0%  -45.15% (p=0.000 n=10)
geomean                                         3.070Ki        1.474Ki       -42.37%                ¹
¹ benchmark set differs from baseline; geomeans may not be comparable

                                              │       1       │                  2                   │
                                              │   allocs/op   │ allocs/op   vs base                  │
SortHashersByValue_Typed_fnv_10-8                 16.000 ± 0%   6.000 ± 0%  -62.50% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8               106.000 ± 0%   6.000 ± 0%  -94.34% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8             1006.000 ± 0%   6.000 ± 0%  -99.40% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8            16.000 ± 0%   6.000 ± 0%  -62.50% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8          106.000 ± 0%   6.000 ± 0%  -94.34% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8         106.000 ± 0%   6.000 ± 0%  -94.34% (p=0.000 n=10)
geomean                                            113.0        6.000       -92.69%                ¹
¹ benchmark set differs from baseline; geomeans may not be comparable
```

Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
2023-06-01 21:41:00 +03:00
c175ef4099 [#8] hrw: Do not create index slice for sorter
`ind` is only needed to index dist or weights, swap them directly.

```
goos: linux
goarch: amd64
pkg: git.frostfs.info/TrueCloudLab/hrw
cpu: 11th Gen Intel(R) Core(TM) i5-1135G7 @ 2.40GHz
                                              │      0      │                  1                   │
                                              │   sec/op    │    sec/op     vs base                │
SortHashersByValue_Typed_fnv_10-8               596.2n ± 4%   580.1n ±  1%   -2.72% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8              4.453µ ± 2%   4.215µ ±  2%   -5.35% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8             41.58µ ± 4%   39.40µ ±  1%   -5.23% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8          624.5n ± 2%   599.6n ±  2%   -3.99% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8         4.593µ ± 2%   4.337µ ±  5%   -5.56% (p=0.003 n=10)
SortHashersByWeightValueTyped_fnv_1000-8        4.896µ ± 8%   4.344µ ±  3%  -11.27% (p=0.000 n=10)
geomean                                         4.668µ        4.400µ         -5.75%

                                              │      0       │                  1                   │
                                              │     B/op     │     B/op      vs base                │
SortHashersByValue_Typed_fnv_10-8                 584.0 ± 0%     472.0 ± 0%  -19.18% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8              4.367Ki ± 0%   3.461Ki ± 0%  -20.75% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8             39.80Ki ± 0%   31.77Ki ± 0%  -20.18% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8            600.0 ± 0%     472.0 ± 0%  -21.33% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8         4.383Ki ± 0%   3.461Ki ± 0%  -21.03% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8        4.383Ki ± 0%   3.461Ki ± 0%  -21.03% (p=0.000 n=10)
geomean                                         3.742Ki        3.070Ki       -17.96%

                                              │      0      │                 1                  │
                                              │  allocs/op  │  allocs/op   vs base               │
SortHashersByValue_Typed_fnv_10-8                17.00 ± 0%    16.00 ± 0%  -5.88% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_100-8               107.0 ± 0%    106.0 ± 0%  -0.93% (p=0.000 n=10)
SortHashersByValue_Typed_fnv_1000-8             1.007k ± 0%   1.006k ± 0%  -0.10% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_10-8           17.00 ± 0%    16.00 ± 0%  -5.88% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_100-8          107.0 ± 0%    106.0 ± 0%  -0.93% (p=0.000 n=10)
SortHashersByWeightValueTyped_fnv_1000-8         107.0 ± 0%    106.0 ± 0%  -0.93% (p=0.000 n=10)
geomean                                          115.3         113.0       -1.94%
```

Signed-off-by: Evgenii Stratonikov <e.stratonikov@yadro.com>
2023-06-01 21:40:46 +03:00
5671632658 [#2] hrw: Add typed methods for hashers
Add generic methods for sort-methods used in node

Signed-off-by: Dmitrii Stepanov <d.stepanov@yadro.com>
2023-02-27 14:18:58 +03:00
22b833d972 Panic if provided value is not a slice
Signed-off-by: Evgenii Stratonikov <stratonikov@runbox.com>
2021-12-28 14:37:17 +03:00
dddcfc8fc5 Simplify SortByValue/Weight a bit
Get rid of unneeded types.
2020-09-11 18:30:18 +03:00
Alex Vanin
f52ea8fb21
Handle NaN values in weight validation function (#7)
NaN values can pass the validation since NaN values don't work
with float arithmetic and comparisons are always true. Now validation
function has check for NaN values.
2019-08-01 12:16:03 +03:00
Alex Vanin
aa230933d1
Move normalization routine out of hrw library (#6)
HRW library supports weighted sorting. Weights must be normalized
before applying. Since there could be different types of normalization
for multiple criteria, there is no point to perform simple
normalization in this library. Pass a slice of normalized weights
to the `SortByWeight` functions.

This commit proposes to:
- remove normalization routine from `SortByWeight` function;
- add `ValidateWeights` function to check if weights are normalized;
- rename `weight` -> `distance` to avoid naming confusion between
  hash distance and actual weights;
- use testify lib in the tests;
2019-07-05 09:49:24 +03:00
Alex Vanin
58a8ce4e47
Added weighted HRW sorting (#5)
* Added weighted HRW sorting

This commit proposes renaming of old `SortByWeight` functions to `Sort`
and implementation of `SortByWeight` function with explicit weights in
arguments. `SortByWeight` function calculates normalized hashes of
nodes and normalized input weights. Then multiplies these values to
obtain node's actual weight for later sorting.

- renamed `SortByWeight` function to `Sort`
- added `SortByWeight`, `SortSliceByWeightValue` and
  `SortSliceBeWeightIndex` functions
- moved code with reflection processing into `prepareRule` function
- added tests and benchmarks for new weighted functions
- added benchmark results into README

* Fixed comments
2019-05-27 10:45:29 +03:00
Evgeniy Kulikov
d32f396204
✌️ Fixed in SortSliceByValue issue when weight calculated twice (#4)
- fixes for SortSliceByValue
- fixes for tests
2019-05-24 12:51:39 +03:00
Evgeniy Kulikov
9a6fcdb1f8
🐌 hot fix types 2019-04-12 14:19:18 +03:00
Evgeniy Kulikov
e06640668d
🐿 Extend SortSliceByValue with more types 2019-04-12 14:09:49 +03:00
Evgeniy Kulikov
e8ba64d915
Small refactoring (#2)
- Use val.Index(0).Interface().(type) in switch-case of SortSliceByValue
- add go1.12 to TravisCI
2019-03-14 15:48:58 +03:00
Alex Vanin
098fd24704 Support for int32 slices (#4) 2019-02-01 12:57:05 +03:00
Evgeniy Kulikov
e5d4b47881
Work around Hashing and tests (#2)
* Work around Hashing and tests

* fixes

* add TravisCI

* fix lint

* so slow

* add badges

* work on README

* Improved tests and get rid of fnv (#3)
2019-01-31 12:54:02 +03:00
Alex Vanin
cf9a7e5b89 Removed redundant m64 variable and tweaked uniform distribution test. (#1) 2019-01-30 12:32:34 +03:00
Evgeniy Kulikov
01e85efd7e
Base HRW implementation in golang 2019-01-30 01:58:30 +03:00