cv4g opened a new pull request, #509:
URL: https://github.com/apache/datasketches-cpp/pull/509
Hi,
thank you so much for considering the other two pull requests. I would like
to propose one more change to `count_min_sketch` focussing on the serialization.
Currently there is a `std::vector` and a `std::ostream&` based API.
ClickHouse is calling `serialize()` to obtain a vector and is then copying it's
contents into another buffer that is then written to disk.
I would now like to remove this additional copy. The ClickHouse project is
trying to avoid using standard library stream mechanisms which is why I would
like to propose to add a sink-based API as an additional option.
Would you be okay with adding this? An advantage is that it could also be
used internally to back both `serialize()` overloads and the `std::ostream&`
overload is benefiting from the change as well.
The change set consists of four commits:
```
Baseline: 6029803 Add micro-benchmark for count_min_sketch serialization
4e66c11 Add serialize_to sink API for count-min sketch
Sink API: 764cdf1 Extend micro-benchmark for count_min_sketch serialization
Final: 63e7f75 Serialize count_min_sketch sketch array in a single
write call
```
4e66c11 introduces a new member function `serialize_to()` and updates both
`serialize()` and `serialize(std::ostream&)` to call it. 63e7f75 improves the
implementation of `serialize_to()` such that it is passing all sketch cells as
a consecutive memory buffer to the sink instead of each one individually.
I've added a microbenchmark once more to ensure it's not pessimising the
existing API:
6029803:
```
Run on (16 X 2000 MHz CPU s)
CPU Caches:
L1 Data 64 KiB (x16)
L1 Instruction 64 KiB (x16)
L2 Unified 2048 KiB (x16)
L3 Unified 36864 KiB (x1)
Load Average: 0.02, 0.07, 0.11
--------------------------------------------------------------------------------------------------
Benchmark Time CPU
Iterations UserCounters...
--------------------------------------------------------------------------------------------------
BM_CountMinGetSerializedSizeBytes/1024 0.482 ns 0.482 ns
1000000000 items_per_second=2.0746G/s
BM_CountMinGetSerializedSizeBytes/4096 0.482 ns 0.482 ns
1000000000 items_per_second=2.07432G/s
BM_CountMinGetSerializedSizeBytes/32768 0.482 ns 0.482 ns
1000000000 items_per_second=2.0739G/s
BM_CountMinGetSerializedSizeBytes/65536 0.482 ns 0.482 ns
1000000000 items_per_second=2.07387G/s
BM_CountMinSerializeVector/1024 1111 ns 1111 ns
628886 bytes_per_second=48.0814G/s
BM_CountMinSerializeVector/4096 4680 ns 4679 ns
149599 bytes_per_second=45.6596G/s
BM_CountMinSerializeVector/32768 52258 ns 52250 ns
12647 bytes_per_second=32.7081G/s
BM_CountMinSerializeVector/65536 119323 ns 119305 ns
5739 bytes_per_second=28.6493G/s
BM_CountMinSerializeOStream/1024 60520 ns 60512 ns
11637 bytes_per_second=904.132M/s
BM_CountMinSerializeOStream/4096 242561 ns 242528 ns
2889 bytes_per_second=902.054M/s
BM_CountMinSerializeOStream/32768 1963452 ns 1963189 ns
356 bytes_per_second=891.418M/s
BM_CountMinSerializeOStream/65536 3928092 ns 3927308 ns
178 bytes_per_second=891.202M/s
```
764cdf1:
```
Load Average: 0.12, 0.09, 0.12
--------------------------------------------------------------------------------------------------
Benchmark Time CPU
Iterations UserCounters...
--------------------------------------------------------------------------------------------------
BM_CountMinGetSerializedSizeBytes/1024 0.482 ns 0.482 ns
1000000000 items_per_second=2.07411G/s
BM_CountMinGetSerializedSizeBytes/4096 0.482 ns 0.482 ns
1000000000 items_per_second=2.07378G/s
BM_CountMinGetSerializedSizeBytes/32768 0.482 ns 0.482 ns
1000000000 items_per_second=2.07327G/s
BM_CountMinGetSerializedSizeBytes/65536 0.482 ns 0.482 ns
1000000000 items_per_second=2.0743G/s
BM_CountMinSerializeVector/1024 15813 ns 15811 ns
44268 bytes_per_second=3.3792G/s
BM_CountMinSerializeVector/4096 63290 ns 63286 ns
11060 bytes_per_second=3.37585G/s
BM_CountMinSerializeVector/32768 507039 ns 506968 ns
1376 bytes_per_second=3.37104G/s
BM_CountMinSerializeVector/65536 1025428 ns 1025387 ns
678 bytes_per_second=3.33337G/s
BM_CountMinSerializeOStream/1024 56433 ns 56425 ns
12404 bytes_per_second=969.608M/s
BM_CountMinSerializeOStream/4096 228901 ns 228888 ns
3059 bytes_per_second=955.807M/s
BM_CountMinSerializeOStream/32768 1863173 ns 1863021 ns
375 bytes_per_second=939.347M/s
BM_CountMinSerializeOStream/65536 3676834 ns 3676281 ns
188 bytes_per_second=952.055M/s
BM_CountMinSerializeToSink/1024 1272 ns 1272 ns
550463 bytes_per_second=42.0145G/s
BM_CountMinSerializeToSink/4096 5159 ns 5158 ns
135711 bytes_per_second=41.4186G/s
BM_CountMinSerializeToSink/32768 43027 ns 43021 ns
16200 bytes_per_second=39.7252G/s
BM_CountMinSerializeToSink/65536 86007 ns 85981 ns
8089 bytes_per_second=39.753G/s
```
63e7f75:
```
Load Average: 0.31, 0.16, 0.14
--------------------------------------------------------------------------------------------------
Benchmark Time CPU
Iterations UserCounters...
--------------------------------------------------------------------------------------------------
BM_CountMinGetSerializedSizeBytes/1024 0.482 ns 0.482 ns
1000000000 items_per_second=2.07442G/s
BM_CountMinGetSerializedSizeBytes/4096 0.482 ns 0.482 ns
1000000000 items_per_second=2.07343G/s
BM_CountMinGetSerializedSizeBytes/32768 0.482 ns 0.482 ns
1000000000 items_per_second=2.07396G/s
BM_CountMinGetSerializedSizeBytes/65536 0.482 ns 0.482 ns
1000000000 items_per_second=2.07333G/s
BM_CountMinSerializeVector/1024 1115 ns 1115 ns
627606 bytes_per_second=47.9082G/s
BM_CountMinSerializeVector/4096 4614 ns 4613 ns
151465 bytes_per_second=46.3141G/s
BM_CountMinSerializeVector/32768 50043 ns 50036 ns
13952 bytes_per_second=34.1558G/s
BM_CountMinSerializeVector/65536 114727 ns 114710 ns
5913 bytes_per_second=29.7968G/s
BM_CountMinSerializeOStream/1024 809 ns 809 ns
867109 bytes_per_second=66.0104G/s
BM_CountMinSerializeOStream/4096 3292 ns 3292 ns
212692 bytes_per_second=64.9082G/s
BM_CountMinSerializeOStream/32768 32502 ns 32501 ns
21627 bytes_per_second=52.5838G/s
BM_CountMinSerializeOStream/65536 76414 ns 76400 ns
9149 bytes_per_second=44.738G/s
BM_CountMinSerializeToSink/1024 710 ns 710 ns
986351 bytes_per_second=75.2844G/s
BM_CountMinSerializeToSink/4096 3189 ns 3189 ns
219511 bytes_per_second=66.9917G/s
BM_CountMinSerializeToSink/32768 32720 ns 32715 ns
21305 bytes_per_second=52.2388G/s
BM_CountMinSerializeToSink/65536 75691 ns 75675 ns
9093 bytes_per_second=45.1667G/s
```
Summary:
```
serialize() vector
Buckets Baseline Sink API Final Final vs Baseline
━━━━━━━━━ ━━━━━━━━━━━ ━━━━━━━━━━━━ ━━━━━━━━━━━ ━━━━━━━━━━━━━━━━━━━
1024 1111 ns 15811 ns 1115 ns 1.00x
───────── ─────────── ──────────── ─────────── ───────────────────
4096 4679 ns 63286 ns 4613 ns 1.01x
───────── ─────────── ──────────── ─────────── ───────────────────
32768 52250 ns 506968 ns 50036 ns 1.04x
───────── ─────────── ──────────── ─────────── ───────────────────
65536 119305 ns 1025387 ns 114710 ns 1.04x
```
The vector serialization regressed badly in 764cdf1 but recovered in 63e7f75
when all sketch cells are written in a single call. It seems that the compiler
was able to optimize the tight loop that copied the sketch cells very well.
```
serialize(std::ostream&)
Buckets Baseline Sink API Final Final vs Baseline
━━━━━━━━━ ━━━━━━━━━━━━ ━━━━━━━━━━━━ ━━━━━━━━━━ ━━━━━━━━━━━━━━━━━━━
1024 60512 ns 56425 ns 809 ns 74.8x
───────── ──────────── ──────────── ────────── ───────────────────
4096 242528 ns 228888 ns 3292 ns 73.7x
───────── ──────────── ──────────── ────────── ───────────────────
32768 1963189 ns 1863021 ns 32501 ns 60.4x
───────── ──────────── ──────────── ────────── ───────────────────
65536 3927308 ns 3676281 ns 76400 ns 51.4x
```
The `ostream&` overload benefits from 63e7f75 a lot.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
