thevar1able/ClickHouse
1
0
Fork 0
mirror of https://github.com/ClickHouse/ClickHouse.git synced 2024-11-19 06:01:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

Continued experiments [#CLICKHOUSE-2]

Browse Source
This commit is contained in:
Alexey Milovidov 2018-01-16 05:04:02 +03:00
parent 2fa8992e2c
commit 846de85799
1 changed files with 10 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
dbms/src/IO/LZ4_decompress_faster.h
View File

@ -9,15 +9,15 @@ namespace LZ4
{ {
/** There are many implementation details of LZ4 decompression loop, that affect performance. /** There are many implementation details of LZ4 decompression loop, that affect performance.
* For example: copy by 8 or by 16 bytes at once; use shuffle instruction to replicate match or not. * For example: copy by 8 or by 16 (SSE2) bytes at once; use shuffle (SSSE3) instruction to replicate match or not.
* *
* The optimal algorithm is dependent: * The optimal algorithm is dependent on:
* *
* - on CPU architecture * 1. CPU architecture.
* (example: on Skylake it's almost always better to copy by 16 bytes and use shuffle, * (example: on Skylake it's almost always better to copy by 16 bytes and use shuffle,
* but on Westmere using shuffle is worse and copy by 16 bytes is better only for high compression ratios) * but on Westmere using shuffle is worse and copy by 16 bytes is better only for high compression ratios)
* *
* - on data distribution * 2. Data distribution.
* (example: when compression ratio is higher than 10.20, * (example: when compression ratio is higher than 10.20,
* it's usually better to copy by 16 bytes rather than 8). * it's usually better to copy by 16 bytes rather than 8).
* *
@ -49,22 +49,22 @@ struct PerformanceStatistics
double count = 0; double count = 0;
double sum = 0; double sum = 0;
double adjusted_count() const double adjustedCount() const
{ {
return count - NUM_INVOCATIONS_TO_THROW_OFF; return count - NUM_INVOCATIONS_TO_THROW_OFF;
} }
double mean() const double mean() const
{ {
return sum / adjusted_count(); return sum / adjustedCount();
} }
/// For better convergence, we don't use proper estimate of stddev. /// For better convergence, we don't use proper estimate of stddev.
/// We want to eventually choose between two algorithms even in case /// We want to eventually separate between two algorithms even in case
/// when there is no statistical significant difference between them. /// when there is no statistical significant difference between them.
double sigma() const double sigma() const
{ {
return mean() / sqrt(adjusted_count()); return mean() / sqrt(adjustedCount());
} }
void update(double seconds, double bytes) void update(double seconds, double bytes)
@ -80,8 +80,8 @@ struct PerformanceStatistics
/// If there is a variant with not enough statistics, always choose it. /// If there is a variant with not enough statistics, always choose it.
/// And in that case prefer variant with less number of invocations. /// And in that case prefer variant with less number of invocations.
if (adjusted_count() < 2) if (adjustedCount() < 2)
return adjusted_count() - 1; return adjustedCount() - 1;
else else
return std::normal_distribution<>(mean(), sigma())(rng); return std::normal_distribution<>(mean(), sigma())(rng);
} }