GH-37293: [C++][Parquet] Encoding: Add Benchmark for DELTA_BYTE_ARRAY by mapleFU · Pull Request #37641 · apache/arrow

mapleFU · 2023-09-09T09:03:19Z

Rationale for this change

Add benchmark for DELTA_BYTE_ARRAY in parquet, and do tiny optimization.

What changes are included in this PR?

Add benchmark for DELTA_BYTE_ARRAY in parquet, and do tiny optimization.

Are these changes tested?

no

Are there any user-facing changes?

no

Closes: [C++][Parquet] Encoding: Add Benchmark for DELTA_BYTE_ARRAY #37293

mapleFU · 2023-09-09T09:03:55Z

- buffer[i].ptr = data_ptr;
- buffer[i].len += prefix_len_ptr[i];
- data_ptr += buffer[i].len;
+ // If the prefix length is zero, the prefix can be ignored.


This is an (maybe-unused) optimization. When prefix == 0, it avoid round of memcpy.

Does it make benchmarks better or worse?

Better, of course, specially when so many prefix == 0

@pitrou I can separate a pr for this. I think we can optimize two cases:
Prefix == 0
Posfix == 0
Each of the case can be well optimize to avoid copying and memory allocation. I can separate for that

github-actions · 2023-09-09T09:04:20Z

⚠️ GitHub issue #37293has been automatically assigned in GitHub to PR creator.

mapleFU · 2023-09-09T09:04:26Z

+ std::vector<ByteArray> values;
+ std::vector<uint8_t> buf(max_length * array_size);
+ values.resize(array_size);
+ prefixed_random_byte_array(array_size, /*seed=*/0, buf.data(), values.data(),


Don't know should we test different prefix configurations
cc @pitrou @wgtmac

CMIW, /*prefixed_probability=*/0.5 is a little bit low which means we can only expect two consecutive string values to share prefixes. It can easily generate a sequence that does not have good prefixes.

Perhaps it would be interesting to also vary prefix length (max_size)?

I think it's enough to vary prefixed_probability. The benchmarking library supports integer parameters, so it can for example be passed as a percentage. Two or three values should be enough (for example 10,90,99).

Ideally, though, we should try to generate data for which the encoding is very space-efficient. Is it the case here?

we should try to generate data for which the encoding is very space-efficient. Is it the case here?
Would 99 percent be "space-efficient"?

Oh I got it, let me mock some tests here.

mapleFU · 2023-09-09T09:07:10Z

cc @pitrou @rok @wgtmac

I've do a basic test, when it's pure random the speed is far away from Plain and other ByteArray Encoding( Though we might get space optimization). I don't know should I test more prefix settings.

wgtmac · 2023-09-11T05:42:51Z

 }
 }
- PARQUET_THROW_NOT_OK(buffered_data_->Resize(data_size));
+ PARQUET_THROW_NOT_OK(buffered_data_->Resize(data_size, false));


Suggested change
PARQUET_THROW_NOT_OK(buffered_data_->Resize(data_size, false));
PARQUET_THROW_NOT_OK(buffered_data_->Resize(data_size, /*shrink_to_fit=*/false));

BTW, we lose the chance to release bulk memory if shrink_to_fit is false. Changing the default behavior of buffered_prefix_length_ above may not make big difference but the actual size of buffered_data_ may vary a lot during the decoding.

Let me leave it first, maybe we need a better method to release memory?

rok · 2023-09-11T11:08:52Z

I've do a basic test, when it's pure random the speed is far away from Plain and other ByteArray Encoding( Though we might get space optimization). I don't know should I test more prefix settings.

Slower encoding is expected, what is the performance difference? 2x, 10x, 100x?

mapleFU · 2023-09-11T11:36:08Z

Slower encoding is expected, what is the performance difference? 2x, 10x, 100x?

Oh my fault. When I use 8-byte and pure random, the others are about 5x. I guess thats mainly because memcpy and more checking. I'll test more ranges this week

pitrou · 2023-09-13T16:15:35Z

Here are some results here.

Decoding doesn't look bad, it's on par with dict decoding (in items/s, I'm not sure the bytes/s figure is consistently computed):

BM_DeltaDecodingByteArray/max-string-length:8/batch-size:512 3444 ns 3443 ns 201688 bytes_per_second=580.339M/s items_per_second=148.719M/s
BM_DeltaDecodingByteArray/max-string-length:64/batch-size:512 3755 ns 3755 ns 183626 bytes_per_second=4.13331G/s items_per_second=136.369M/s
BM_DeltaDecodingByteArray/max-string-length:1024/batch-size:512 5380 ns 5380 ns 129118 bytes_per_second=47.3178G/s items_per_second=95.1757M/s
BM_DeltaDecodingByteArray/max-string-length:8/batch-size:2048 14293 ns 14291 ns 48895 bytes_per_second=550.606M/s items_per_second=143.306M/s
BM_DeltaDecodingByteArray/max-string-length:64/batch-size:2048 15597 ns 15595 ns 44892 bytes_per_second=3.93516G/s items_per_second=131.323M/s
BM_DeltaDecodingByteArray/max-string-length:1024/batch-size:2048 23939 ns 23934 ns 29173 bytes_per_second=41.7182G/s items_per_second=85.5671M/s
BM_DictDecodingByteArray/max-string-length:8/batch-size:512 2953 ns 2952 ns 234887 bytes_per_second=1.29091G/s items_per_second=173.433M/s
BM_DictDecodingByteArray/max-string-length:64/batch-size:512 4158 ns 4157 ns 169321 bytes_per_second=4.13726G/s items_per_second=123.171M/s
BM_DictDecodingByteArray/max-string-length:1024/batch-size:512 10668 ns 10666 ns 65659 bytes_per_second=23.0956G/s items_per_second=48.0035M/s
BM_DictDecodingByteArray/max-string-length:8/batch-size:2048 8789 ns 8788 ns 75111 bytes_per_second=1.7316G/s items_per_second=233.051M/s
BM_DictDecodingByteArray/max-string-length:64/batch-size:2048 13147 ns 13144 ns 54072 bytes_per_second=5.20906G/s items_per_second=155.814M/s
BM_DictDecodingByteArray/max-string-length:1024/batch-size:2048 46555 ns 46544 ns 15010 bytes_per_second=21.0795G/s items_per_second=44.0014M/s

mapleFU · 2023-09-15T11:44:56Z

BM_DeltaEncodingByteArray/min-prefix-string-length:0/max-string-length:8/batch-size:512/prefixed-probability:10 11626 ns 11623 ns 59332 bytes_per_second=175.745M/s items_per_second=44.0488M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:4/max-string-length:8/batch-size:512/prefixed-probability:90 12596 ns 12588 ns 55529 bytes_per_second=233.718M/s items_per_second=40.6731M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:8/max-string-length:8/batch-size:512/prefixed-probability:99 11033 ns 11028 ns 63642 bytes_per_second=354.205M/s items_per_second=46.4263M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:0/max-string-length:8/batch-size:2048/prefixed-probability:10 51175 ns 51159 ns 13558 bytes_per_second=156.121M/s items_per_second=40.0319M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:4/max-string-length:8/batch-size:2048/prefixed-probability:90 56700 ns 56422 ns 12403 bytes_per_second=207.292M/s items_per_second=36.2977M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:8/max-string-length:8/batch-size:2048/prefixed-probability:99 46428 ns 46191 ns 15185 bytes_per_second=338.272M/s items_per_second=44.338M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:0/max-string-length:64/batch-size:512/prefixed-probability:10 21084 ns 21078 ns 33275 bytes_per_second=728.956M/s items_per_second=24.2912M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:32/max-string-length:64/batch-size:512/prefixed-probability:90 23528 ns 23403 ns 28746 bytes_per_second=1006.64M/s items_per_second=21.8774M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:64/max-string-length:64/batch-size:512/prefixed-probability:99 20969 ns 20863 ns 33702 bytes_per_second=1.46274G/s items_per_second=24.5407M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:0/max-string-length:64/batch-size:2048/prefixed-probability:10 114492 ns 112077 ns 6082 bytes_per_second=567.938M/s items_per_second=18.2731M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:32/max-string-length:64/batch-size:2048/prefixed-probability:90 118275 ns 117342 ns 5873 bytes_per_second=805.154M/s items_per_second=17.4532M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:64/max-string-length:64/batch-size:2048/prefixed-probability:99 116968 ns 115977 ns 6185 bytes_per_second=1077.8M/s items_per_second=17.6587M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:0/max-string-length:1024/batch-size:512/prefixed-probability:10 133510 ns 132742 ns 5261 bytes_per_second=1.90002G/s items_per_second=3.85712M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:512/max-string-length:1024/batch-size:512/prefixed-probability:90 191934 ns 191903 ns 3513 bytes_per_second=1.91668G/s items_per_second=2.66801M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:1024/max-string-length:1024/batch-size:512/prefixed-probability:99 272684 ns 272636 ns 2565 bytes_per_second=1.79096G/s items_per_second=1.87796M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:0/max-string-length:1024/batch-size:2048/prefixed-probability:10 622002 ns 621948 ns 1193 bytes_per_second=1.59998G/s items_per_second=3.29288M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:512/max-string-length:1024/batch-size:2048/prefixed-probability:90 799290 ns 799220 ns 870 bytes_per_second=1.8367G/s items_per_second=2.5625M/s
BM_DeltaEncodingByteArray/min-prefix-string-length:1024/max-string-length:1024/batch-size:2048/prefixed-probability:99 1176743 ns 1176583 ns 628 bytes_per_second=1.66G/s items_per_second=1.74063M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:0/max-string-length:8/batch-size:512/prefixed-probability:10 2387 ns 2371 ns 295677 bytes_per_second=861.51M/s items_per_second=215.929M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:4/max-string-length:8/batch-size:512/prefixed-probability:90 2452 ns 2452 ns 284102 bytes_per_second=1.17173G/s items_per_second=208.805M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:8/max-string-length:8/batch-size:512/prefixed-probability:99 2417 ns 2417 ns 289162 bytes_per_second=1.57817G/s items_per_second=211.818M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:0/max-string-length:8/batch-size:2048/prefixed-probability:10 8744 ns 8743 ns 79466 bytes_per_second=913.521M/s items_per_second=234.241M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:4/max-string-length:8/batch-size:2048/prefixed-probability:90 8962 ns 8958 ns 78023 bytes_per_second=1.27509G/s items_per_second=228.633M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:8/max-string-length:8/batch-size:2048/prefixed-probability:99 8937 ns 8936 ns 78077 bytes_per_second=1.70758G/s items_per_second=229.188M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:0/max-string-length:64/batch-size:512/prefixed-probability:10 2428 ns 2427 ns 287872 bytes_per_second=6.18231G/s items_per_second=210.959M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:32/max-string-length:64/batch-size:512/prefixed-probability:90 2410 ns 2410 ns 288149 bytes_per_second=9.54651G/s items_per_second=212.454M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:64/max-string-length:64/batch-size:512/prefixed-probability:99 2460 ns 2460 ns 284264 bytes_per_second=12.4073G/s items_per_second=208.16M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:0/max-string-length:64/batch-size:2048/prefixed-probability:10 9021 ns 9020 ns 77622 bytes_per_second=6.89111G/s items_per_second=227.039M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:32/max-string-length:64/batch-size:2048/prefixed-probability:90 8871 ns 8870 ns 78905 bytes_per_second=10.4019G/s items_per_second=230.892M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:64/max-string-length:64/batch-size:2048/prefixed-probability:99 9176 ns 9176 ns 76328 bytes_per_second=13.3033G/s items_per_second=223.193M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:0/max-string-length:1024/batch-size:512/prefixed-probability:10 2504 ns 2504 ns 278691 bytes_per_second=100.718G/s items_per_second=204.462M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:512/max-string-length:1024/batch-size:512/prefixed-probability:90 2525 ns 2525 ns 277716 bytes_per_second=145.692G/s items_per_second=202.803M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:1024/max-string-length:1024/batch-size:512/prefixed-probability:99 2558 ns 2556 ns 273213 bytes_per_second=190.996G/s items_per_second=200.274M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:0/max-string-length:1024/batch-size:2048/prefixed-probability:10 9260 ns 9247 ns 75385 bytes_per_second=107.617G/s items_per_second=221.484M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:512/max-string-length:1024/batch-size:2048/prefixed-probability:90 9304 ns 9297 ns 74922 bytes_per_second=157.898G/s items_per_second=220.293M/s
BM_DeltaDecodingByteArray/min-prefix-string-length:1024/max-string-length:1024/batch-size:2048/prefixed-probability:99 9000 ns 8998 ns 77475 bytes_per_second=217.05G/s items_per_second=227.594M/s

@pitrou I've test the Decode/Encoding on MacOS, to avoid too much benchmarks, I configure the min-prefix-length in testing.

When string is nearly pure randomly(prefix-prob = 10%, min-prefix-length = 0), because of my optimization in this patch, it could be a bit slowly than Plain.
When string is partial randomly (prefix-prob = 50%, min-prefix-length = max-length/2), decode/encode will consuming some time
When string is highly repeated (prefix-prob = 99%, min-prefix-length = max-length), decode/encode will become fast again

mapleFU · 2023-09-15T19:10:01Z

Comparing with Plain Decoding and Delta length, Delta would be slower, because Plain/Delta length would not need to copy any data. Their speed is stable with different input.

BM_PlainDecodingByteArray/max-string-length:8/batch-size:512 1279 ns 1279 ns 548306 bytes_per_second=2.97226G/s items_per_second=400.298M/s
BM_PlainDecodingByteArray/max-string-length:64/batch-size:512 1279 ns 1278 ns 543322 bytes_per_second=13.5389G/s items_per_second=400.49M/s
BM_PlainDecodingByteArray/max-string-length:1024/batch-size:512 1278 ns 1278 ns 547482 bytes_per_second=198.153G/s items_per_second=400.56M/s

BM_DeltaLengthDecodingByteArray/max-string-length:8/batch-size:512 1566 ns 1566 ns 449207 bytes_per_second=2.42814G/s items_per_second=327.017M/s
BM_DeltaLengthDecodingByteArray/max-string-length:64/batch-size:512 1576 ns 1576 ns 450990 bytes_per_second=10.9846G/s items_per_second=324.93M/s
BM_DeltaLengthDecodingByteArray/max-string-length:1024/batch-size:512 1554 ns 1554 ns 447002 bytes_per_second=162.973G/s items_per_second=329.444M/s

Comparing with Dict, Dict require to set dictionary, and parsing dictionary would get a memcpy. So in benchmark, they're similiar. However, during real using Dict, a ColumnChunk will only set dictionary once, so Dictionary is still faster than DELTA.

Currently DELTA doesn't has speed advantage. However, if data is prefixed-data, like "abcd, abcf, ...", DELTA can balance space and encoding. And when real using parquet, user usally do "compression + encoding", though DELTA decode might be a bit slower, but the decompression part would be so fast because data is short. And I think user can use it if he find the pattern is suitable, user should not use it by default.

cc @rok

rok · 2023-09-15T19:32:21Z

@mapleFU Agreed! Users should be aware why and when to use which encoding.
One of the use cases for DELTA is storing sorted arrays of file paths.

mapleFU · 2023-09-15T20:34:50Z

Yeah, just decoding/encoding cannot prove that the ability of delta(Though it's still important)

I've test encoding with compression and DELTA gets ok

mapleFU · 2023-09-19T14:40:50Z

@pitrou I've update the benchmark, would you mind take a look?

pitrou · 2023-09-26T12:39:46Z

 // all the prefix lengths are buffered in buffered_prefix_length_.
- PARQUET_THROW_NOT_OK(buffered_prefix_length_->Resize(num_prefix * sizeof(int32_t)));
+ PARQUET_THROW_NOT_OK(buffered_prefix_length_->Resize(num_prefix * sizeof(int32_t),
+ /*shrink_to_fit=*/false));


Is this desirable? If decoding first a large page and then a small page, this means that memory wouldn't be released. Does page size always stay similar?

Emmm Personally I don't think the page will be too large, but user might set large page-size, I'll remove that

mapleFU · 2023-09-26T13:15:20Z

if (buffer[i].len == 0) {
// If the buffer length is zero, the memcpy to data_ptr can be ignored.
buffer[i].ptr = reinterpret_cast<constuint8_t*>(prefix.data());
buffer[i].len = prefix_len_ptr[i];
} elseif (prefix_len_ptr[i] != 0) {
// If the prefix length is zero, the prefix can be ignored.memcpy(data_ptr, prefix.data(), prefix_len_ptr[i]);
// buffer[i] currently points to the string suffixmemcpy(data_ptr + prefix_len_ptr[i], buffer[i].ptr, buffer[i].len);
buffer[i].ptr = data_ptr;
buffer[i].len += prefix_len_ptr[i];
data_ptr += buffer[i].len;
}
prefix = std::string_view{buffer[i]};

I tent to use code like this, but let me leave it to later patch with more carefully design...

pitrou · 2023-09-26T13:39:36Z

I improved and simplified the benchmark code. Here are the current results here:

BM_DeltaEncodingByteArray/max-string-length:8/batch-size:512/prefixed-percent:10 11142 ns 11140 ns 62716 bytes_per_second=175.236M/s compression_ratio=0.700959 items_per_second=45.9596M/s
BM_DeltaEncodingByteArray/max-string-length:8/batch-size:512/prefixed-percent:90 11904 ns 11902 ns 58803 bytes_per_second=172.829M/s compression_ratio=0.765799 items_per_second=43.0166M/s
BM_DeltaEncodingByteArray/max-string-length:8/batch-size:512/prefixed-percent:99 11796 ns 11794 ns 58970 bytes_per_second=169.732M/s compression_ratio=0.759524 items_per_second=43.4132M/s
BM_DeltaEncodingByteArray/max-string-length:8/batch-size:2048/prefixed-percent:10 40895 ns 40885 ns 16988 bytes_per_second=186.932M/s compression_ratio=0.695357 items_per_second=50.0914M/s
BM_DeltaEncodingByteArray/max-string-length:8/batch-size:2048/prefixed-percent:90 45275 ns 45266 ns 15099 bytes_per_second=172.528M/s compression_ratio=0.753808 items_per_second=45.2437M/s
BM_DeltaEncodingByteArray/max-string-length:8/batch-size:2048/prefixed-percent:99 46609 ns 46598 ns 14949 bytes_per_second=168.19M/s compression_ratio=0.762051 items_per_second=43.9504M/s
BM_DeltaEncodingByteArray/max-string-length:64/batch-size:512/prefixed-percent:10 14330 ns 14326 ns 48859 bytes_per_second=1085.98M/s compression_ratio=0.933408 items_per_second=35.7382M/s
BM_DeltaEncodingByteArray/max-string-length:64/batch-size:512/prefixed-percent:90 14291 ns 14288 ns 49084 bytes_per_second=1104.8M/s compression_ratio=1.19315 items_per_second=35.8346M/s
BM_DeltaEncodingByteArray/max-string-length:64/batch-size:512/prefixed-percent:99 14001 ns 13998 ns 49811 bytes_per_second=1100.33M/s compression_ratio=1.21214 items_per_second=36.5759M/s
BM_DeltaEncodingByteArray/max-string-length:64/batch-size:2048/prefixed-percent:10 46585 ns 46575 ns 14949 bytes_per_second=1.29997G/s compression_ratio=0.930448 items_per_second=43.9721M/s
BM_DeltaEncodingByteArray/max-string-length:64/batch-size:2048/prefixed-percent:90 60950 ns 60938 ns 11346 bytes_per_second=1028.95M/s compression_ratio=1.18361 items_per_second=33.6078M/s
BM_DeltaEncodingByteArray/max-string-length:64/batch-size:2048/prefixed-percent:99 60273 ns 60260 ns 11491 bytes_per_second=1040.17M/s compression_ratio=1.22369 items_per_second=33.9862M/s
BM_DeltaEncodingByteArray/max-string-length:1024/batch-size:512/prefixed-percent:10 40006 ns 39997 ns 17513 bytes_per_second=5.95053G/s compression_ratio=1.0309 items_per_second=12.8011M/s
BM_DeltaEncodingByteArray/max-string-length:1024/batch-size:512/prefixed-percent:90 54380 ns 54368 ns 12895 bytes_per_second=4.32286G/s compression_ratio=1.43071 items_per_second=9.41729M/s
BM_DeltaEncodingByteArray/max-string-length:1024/batch-size:512/prefixed-percent:99 59008 ns 58992 ns 11894 bytes_per_second=4.28957G/s compression_ratio=1.54352 items_per_second=8.67907M/s
BM_DeltaEncodingByteArray/max-string-length:1024/batch-size:2048/prefixed-percent:10 150360 ns 150330 ns 4295 bytes_per_second=6.41765G/s compression_ratio=1.02877 items_per_second=13.6234M/s
BM_DeltaEncodingByteArray/max-string-length:1024/batch-size:2048/prefixed-percent:90 189980 ns 189936 ns 3689 bytes_per_second=5.15671G/s compression_ratio=1.41325 items_per_second=10.7826M/s
BM_DeltaEncodingByteArray/max-string-length:1024/batch-size:2048/prefixed-percent:99 203198 ns 203155 ns 3417 bytes_per_second=4.93974G/s compression_ratio=1.50832 items_per_second=10.081M/s
BM_DeltaDecodingByteArray/max-string-length:8/batch-size:512/prefixed-percent:10 5270 ns 5270 ns 132014 bytes_per_second=370.447M/s compression_ratio=0.700959 items_per_second=97.1578M/s
BM_DeltaDecodingByteArray/max-string-length:8/batch-size:512/prefixed-percent:90 5061 ns 5060 ns 126446 bytes_per_second=406.504M/s compression_ratio=0.765799 items_per_second=101.178M/s
BM_DeltaDecodingByteArray/max-string-length:8/batch-size:512/prefixed-percent:99 5542 ns 5542 ns 126490 bytes_per_second=361.218M/s compression_ratio=0.759524 items_per_second=92.3904M/s
BM_DeltaDecodingByteArray/max-string-length:8/batch-size:2048/prefixed-percent:10 23027 ns 23024 ns 30212 bytes_per_second=331.951M/s compression_ratio=0.695357 items_per_second=88.9516M/s
BM_DeltaDecodingByteArray/max-string-length:8/batch-size:2048/prefixed-percent:90 22239 ns 22235 ns 31329 bytes_per_second=351.231M/s compression_ratio=0.753808 items_per_second=92.1068M/s
BM_DeltaDecodingByteArray/max-string-length:8/batch-size:2048/prefixed-percent:99 22392 ns 22389 ns 31045 bytes_per_second=350.055M/s compression_ratio=0.762051 items_per_second=91.4744M/s
BM_DeltaDecodingByteArray/max-string-length:64/batch-size:512/prefixed-percent:10 5040 ns 5040 ns 137874 bytes_per_second=3.01475G/s compression_ratio=0.933408 items_per_second=101.592M/s
BM_DeltaDecodingByteArray/max-string-length:64/batch-size:512/prefixed-percent:90 5312 ns 5312 ns 133927 bytes_per_second=2.90216G/s compression_ratio=1.19315 items_per_second=96.3918M/s
BM_DeltaDecodingByteArray/max-string-length:64/batch-size:512/prefixed-percent:99 5238 ns 5238 ns 131598 bytes_per_second=2.87179G/s compression_ratio=1.21214 items_per_second=97.7513M/s
BM_DeltaDecodingByteArray/max-string-length:64/batch-size:2048/prefixed-percent:10 20561 ns 20558 ns 33335 bytes_per_second=2.9452G/s compression_ratio=0.930448 items_per_second=99.6227M/s
BM_DeltaDecodingByteArray/max-string-length:64/batch-size:2048/prefixed-percent:90 23385 ns 23381 ns 28702 bytes_per_second=2.61889G/s compression_ratio=1.18361 items_per_second=87.5921M/s
BM_DeltaDecodingByteArray/max-string-length:64/batch-size:2048/prefixed-percent:99 22930 ns 22925 ns 29705 bytes_per_second=2.67001G/s compression_ratio=1.22369 items_per_second=89.333M/s
BM_DeltaDecodingByteArray/max-string-length:1024/batch-size:512/prefixed-percent:10 8253 ns 8251 ns 85126 bytes_per_second=28.8434G/s compression_ratio=1.0309 items_per_second=62.0494M/s
BM_DeltaDecodingByteArray/max-string-length:1024/batch-size:512/prefixed-percent:90 9089 ns 9088 ns 76982 bytes_per_second=25.8625G/s compression_ratio=1.43071 items_per_second=56.341M/s
BM_DeltaDecodingByteArray/max-string-length:1024/batch-size:512/prefixed-percent:99 9418 ns 9416 ns 74093 bytes_per_second=26.874G/s compression_ratio=1.54352 items_per_second=54.374M/s
BM_DeltaDecodingByteArray/max-string-length:1024/batch-size:2048/prefixed-percent:10 35156 ns 35151 ns 19918 bytes_per_second=27.4465G/s compression_ratio=1.02877 items_per_second=58.2635M/s
BM_DeltaDecodingByteArray/max-string-length:1024/batch-size:2048/prefixed-percent:90 38266 ns 38257 ns 18211 bytes_per_second=25.6015G/s compression_ratio=1.41325 items_per_second=53.532M/s
BM_DeltaDecodingByteArray/max-string-length:1024/batch-size:2048/prefixed-percent:99 38955 ns 38948 ns 17954 bytes_per_second=25.7662G/s compression_ratio=1.50832 items_per_second=52.5835M/s

mapleFU · 2023-09-26T13:45:46Z

Nice! So currently the memcpy will causing high CPU usage, let me optimize with #37641 (comment) after this is merged, thanks!

pitrou

+1, let's wait for CI

mapleFU · 2023-09-26T14:21:06Z

Ooops, I believe these failed CI is not caused by me...

pitrou · 2023-09-26T14:22:21Z

Yes, the CI failures are unrelated.

conbench-apache-arrow · 2023-09-27T15:24:18Z

After merging your PR, Conbench analyzed the 6 benchmarking runs that have been run so far on merge-commit 5978729.

There were no benchmark performance regressions. 🎉

The full Conbench report has more details. It also includes information about possible false positives for unstable benchmarks that are known to sometimes produce them.

…_ARRAY (apache#37641) ### Rationale for this change Add benchmark for DELTA_BYTE_ARRAY in parquet, and do tiny optimization. ### What changes are included in this PR? Add benchmark for DELTA_BYTE_ARRAY in parquet, and do tiny optimization. ### Are these changes tested? no ### Are there any user-facing changes? no * Closes: apache#37293 Lead-authored-by: mwish <maplewish117@gmail.com> Co-authored-by: Antoine Pitrou <antoine@python.org> Signed-off-by: Antoine Pitrou <antoine@python.org>

add benchmark for delta
c99d49d

mapleFU requested a review from wgtmac as a code owner September 9, 2023 09:03

mapleFU commented Sep 9, 2023
View reviewed changes

github-actionsBot added the awaiting review Awaiting review label Sep 9, 2023

github-actionsBot added Component: Parquet Component: C++ labels Sep 9, 2023

mapleFU commented Sep 9, 2023
View reviewed changes

github-actionsBot added awaiting committer review Awaiting committer review and removed awaiting review Awaiting review labels Sep 9, 2023

wgtmac reviewed Sep 11, 2023
View reviewed changes

github-actionsBot added awaiting changes Awaiting changes and removed awaiting committer review Awaiting committer review labels Sep 11, 2023

mapleFU added 2 commits September 15, 2023 18:44

Merge branch 'main' into parquet-encoding-benchmark-delta
4feb5c0

fix comment
d9825d2

github-actionsBot added awaiting change review Awaiting change review and removed awaiting changes Awaiting changes labels Sep 15, 2023

mapleFU force-pushed the parquet-encoding-benchmark-delta branch from c2bde25 to 971e2c9Compare September 15, 2023 11:58

refactor test
c567ed6

mapleFU force-pushed the parquet-encoding-benchmark-delta branch from 971e2c9 to c567ed6Compare September 15, 2023 12:30

mapleFU requested review from pitrou and rok September 19, 2023 05:37

mapleFU requested a review from wgtmac September 19, 2023 05:37

pitrou reviewed Sep 26, 2023
View reviewed changes

mapleFU added 2 commits September 26, 2023 21:01

Merge branch 'main' into parquet-encoding-benchmark-delta
9a8d100

resolve comment and revert optimize
e8e8766

Simplify and improve benchmarks
557d0ca

pitrou approved these changes Sep 26, 2023
View reviewed changes

pitrou merged commit 5978729 into apache:mainSep 26, 2023

pitrou removed the awaiting change review Awaiting change review label Sep 26, 2023

mapleFU deleted the parquet-encoding-benchmark-delta branch September 26, 2023 14:23

	PARQUET_THROW_NOT_OK(buffered_data_->Resize(data_size, false));
	PARQUET_THROW_NOT_OK(buffered_data_->Resize(data_size, /shrink_to_fit=/false));

Uh oh!

Conversation

mapleFU commented Sep 9, 2023• edited by github-actions BotLoading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Rationale for this change

What changes are included in this PR?

Are these changes tested?

Are there any user-facing changes?

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mapleFUSep 26, 2023• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

github-actionsBot commented Sep 9, 2023

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rokSep 11, 2023• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mapleFU commented Sep 9, 2023• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

rok commented Sep 11, 2023

Uh oh!

mapleFU commented Sep 11, 2023

Uh oh!

pitrou commented Sep 13, 2023

Uh oh!

mapleFU commented Sep 15, 2023• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mapleFU commented Sep 15, 2023• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rok commented Sep 15, 2023

Uh oh!

mapleFU commented Sep 15, 2023

Uh oh!

mapleFU commented Sep 19, 2023

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mapleFU commented Sep 26, 2023

Uh oh!

pitrou commented Sep 26, 2023

Uh oh!

mapleFU commented Sep 26, 2023

Uh oh!

pitrou left a comment

Choose a reason for hiding this comment

Uh oh!

mapleFU commented Sep 26, 2023

Uh oh!

mapleFU commented Sep 9, 2023•
edited by github-actions Bot
Loading

mapleFUSep 26, 2023•
edited
Loading

rokSep 11, 2023•
edited
Loading

mapleFU commented Sep 9, 2023•
edited
Loading

mapleFU commented Sep 15, 2023•
edited
Loading

mapleFU commented Sep 15, 2023•
edited
Loading