震精 - PostgreSQL decimal64 decimal128 高效率数值 类型扩展
标签
PostgreSQL , decimal64 , decimal128 , float4 , float8 , numeric
背景
PostgreSQL内置的数值类型包括
整型、浮点、整型序列、"无限"精度数值
| Name | Storage Size | Description | Range |
|---|---|---|---|
| smallint | 2 bytes | small-range integer | -32768 to +32767 |
| integer | 4 bytes | typical choice for integer | -2147483648 to +2147483647 |
| bigint | 8 bytes | large-range integer | -9223372036854775808 to +9223372036854775807 |
| decimal | variable | user-specified precision, exact | up to 131072 digits before the decimal point; up to 16383 digits after the decimal point |
| numeric | variable | user-specified precision, exact | up to 131072 digits before the decimal point; up to 16383 digits after the decimal point |
| real | 4 bytes | variable-precision, inexact | 6 decimal digits precision |
| double precision | 8 bytes | variable-precision, inexact | 15 decimal digits precision |
| smallserial | 2 bytes | small autoincrementing integer | 1 to 32767 |
| serial | 4 bytes | autoincrementing integer | 1 to 2147483647 |
| bigserial | 8 bytes | large autoincrementing integer | 1 to 9223372036854775807 |
其中除了 "无限"精度数值类型。他类型都是定长存储，使用时不需要调用palloc，效率较高。
如果你要使用超过双精能表示的有效范围的数值，目前只能选择decimal\numeric类型，而这个类型前面说了，由于是变长设计，需要调用palloc，效率一般。
那么在数据分析领域，或者需要处理非常多的数据记录时，numeric类型的开销是较大的。
PostgreSQL社区有一些扩展，可以解决这个问题，
1. 比如2nd的fixeddecimal插件，使用INT8来表示NUMERIC，精度可调。
《PostgreSQL fixeddecimal - 用CPU "硬解码" 提升1倍 数值运算能力 助力金融大数据量计算》
2. 比如社区的pgdecimal插件，支持decimal32和decimal64两种类型。
https://pgxn.org/dist/pgdecimal/1.0.0/
3. 比如vitesse的pgdecimal插件，也就是本文将提到的插件，支持decimal64与decimal128类型，精度基本上足够使用。
推荐使用vitesse提供的pgdecimal插件，因为它效率够高，精度够大。
pgdecimal插件介绍
有两个常见的decimal库，decNumber以及Intel提供的Intel ADX库。
pgdecimal插件选择了decNumber库，因为GCC也在用它（法律风险更小？）
https://github.com/gcc-mirror/gcc/tree/master/libdecnumber
decimal库的性能对比
http://speleotrove.com/decimal/dpintro.html
decNumber与Inter ADX性能接近，但是Inter ADX提供了decimal64/128, int32/64, float/double类型的相互转换，这个很给力。（也许将来vitesse会支持intel adx库吧）
pgdecimal 依赖的decNumber，因此我们必须先安装decNumber
decNumber安装
1. 下载 decNumber package
http://speleotrove.com/decimal/
wget http://speleotrove.com/decimal/decNumber-icu-368.zip
unzip decNumber-icu-368.zip
或者从本站链接下载
2. 安装decNumber到postgresql软件目录中(假设postgresql安装在/home/digoal/pgsql9.6)
首先要在postgresql软件的include目录中，创建一个空目录，
mkdir -p /home/digoal/pgsql9.6/include/decnumber
在decNumber src目录中创建Makefile，install -D 修改为对应要安装的目录。
cd decNumber
vi Makefile
OBJS = decSingle.o decDouble.o decQuad.o decNumber.o decContext.o
CFLAGS = -Wall -g -O2 -fPIC
libdecnumber.a: $(OBJS)
ar -rcs libdecnumber.a $(OBJS)
clean:
rm -f libdecnumber.a $(OBJS)
install:
install -D *.h /home/digoal/pgsql9.6/include/decnumber
install -D libdecnumber.a /home/digoal/pgsql9.6/lib
3. 编译安装decNumber
cd decNumber
make
make install
4. decNumber的C库reference如下, pgdecimal插件中用到的decnumber库，需要了解细节的话请参考:
The decNumber C library
pgdecimal安装
git clone https://github.com/vitesse-ftian/pgdecimal
或者从本站链接下载
cd pgdecimal
有一个小BUG，.control的版本号没有与sql文件的版本号对齐
mv decimal--2.0.sql decimal--1.0.sql
另外，需要修改一下Makefile，指定版本，以及decnumber的include和lib目录
vi Makefile
PG_CPPFLAGS = -I/home/digoal/pgsql9.6/include/decnumber
SHLIB_LINK = -L/home/digoal/pgsql9.6/lib -ldecnumber
DATA = decimal--1.0.sql
安装
export PATH=/home/digoal/pgsql9.6/bin:$PATH
USE_PGXS=1 make clean
USE_PGXS=1 make
USE_PGXS=1 make install
/bin/mkdir -p '/home/digoal/pgsql9.6/lib'
/bin/mkdir -p '/home/digoal/pgsql9.6/share/extension'
/bin/mkdir -p '/home/digoal/pgsql9.6/share/extension'
/usr/bin/install -c -m 755 decimal.so '/home/digoal/pgsql9.6/lib/decimal.so'
/usr/bin/install -c -m 644 .//decimal.control '/home/digoal/pgsql9.6/share/extension/'
/usr/bin/install -c -m 644 .//decimal--1.0.sql '/home/digoal/pgsql9.6/share/extension/'
使用
psql
postgres=# create extension decimal;
CREATE EXTENSION
pgdecimal性能对比
使用int8, float8, decimal64, decimal128, numeric(15,3) 几种类型，分别比较这几种类型的性能。
create table tt(ii bigint, d double precision, d64 decimal64, d128 decimal128, n numeric(15, 3));
postgres=# \timing
Timing is on.
生成测试数据
postgres=# insert into tt select i, i + 0.123, i + 0.123::decimal64, i + 0.123::decimal128, i + 0.123 from generate_series(1, 1000000) i;
INSERT 0 1000000
Time: 2125.723 ms
postgres=# select * from tt limit 2;
ii | d | d64 | d128 | n
----+-------+-------+-------+-------
1 | 1.123 | 1.123 | 1.123 | 1.123
2 | 2.123 | 2.123 | 2.123 | 2.123
(2 rows)
普通查询性能对比
postgres=# select count(*) from tt where (d + d*d + d*d*d + d*d*d*d) > 10000000;
count
--------
999945
(1 row)
Time: 411.418 ms
postgres=# select count(*) from tt where (n + n*n + n*n*n + n*n*n*n) > 10000000;
count
--------
999945
(1 row)
Time: 1949.367 ms
postgres=# select count(*) from tt where (d64 + d64*d64 + d64*d64*d64 + d64*d64*d64*d64) > 10000000;
count
--------
999945
(1 row)
Time: 1165.304 ms
postgres=# select count(*) from tt where (d128 + d128*d128 + d128*d128*d128 + d128*d128*d128*d128) > 10000000;
count
--------
999945
(1 row)
Time: 1517.179 ms
排序性能对比
postgres=# select * from tt order by d limit 2 offset 999000;
ii | d | d64 | d128 | n
--------+------------+------------+------------+------------
999001 | 999001.123 | 999001.123 | 999001.123 | 999001.123
999002 | 999002.123 | 999002.123 | 999002.123 | 999002.123
(2 rows)
Time: 804.645 ms
postgres=# select * from tt order by n limit 2 offset 999000;
ii | d | d64 | d128 | n
--------+------------+------------+------------+------------
999001 | 999001.123 | 999001.123 | 999001.123 | 999001.123
999002 | 999002.123 | 999002.123 | 999002.123 | 999002.123
(2 rows)
Time: 2828.066 ms
postgres=# select * from tt order by d64 limit 2 offset 999000;
ii | d | d64 | d128 | n
--------+------------+------------+------------+------------
999001 | 999001.123 | 999001.123 | 999001.123 | 999001.123
999002 | 999002.123 | 999002.123 | 999002.123 | 999002.123
(2 rows)
Time: 1826.044 ms
postgres=# select * from tt order by d128 limit 2 offset 999000;
ii | d | d64 | d128 | n
--------+------------+------------+------------+------------
999001 | 999001.123 | 999001.123 | 999001.123 | 999001.123
999002 | 999002.123 | 999002.123 | 999002.123 | 999002.123
(2 rows)
Time: 2118.647 ms
哈希JOIN性能对比
postgres=# explain select count(*) from tt t1 join tt t2 on t1.d64 * t1.d64 + t1.d64 = t2.d64 + t2.d64 * t2.d64;
QUERY PLAN
----------------------------------------------------------------------------------
Aggregate (cost=6875071228.00..6875071228.01 rows=1 width=8)
-> Hash Join (cost=36707.00..5625071228.00 rows=500000000000 width=0)
Hash Cond: (((t1.d64 * t1.d64) + t1.d64) = (t2.d64 + (t2.d64 * t2.d64)))
-> Seq Scan on tt t1 (cost=0.00..20300.00 rows=1000000 width=8)
-> Hash (cost=20300.00..20300.00 rows=1000000 width=8)
-> Seq Scan on tt t2 (cost=0.00..20300.00 rows=1000000 width=8)
(6 rows)
Time: 0.508 ms
postgres=# select count(*) from tt t1 join tt t2 on t1.d64 * t1.d64 + t1.d64 = t2.d64 + t2.d64 * t2.d64;
count
---------
1000000
(1 row)
Time: 1681.451 ms
postgres=# select count(*) from tt t1 join tt t2 on t1.n * t1.n + t1.n = t2.n + t2.n * t2.n;
count
---------
1000000
(1 row)
Time: 2395.894 ms
嵌套循环性能对比
postgres=# explain select count(*) from tt t1, tt t2 where t1.ii < 10000 and t2.ii < 10000 and t1.d * t1.d + t1.d > t2.d + t2.d * t2.d;
QUERY PLAN
-------------------------------------------------------------------------------
Aggregate (cost=2699703.15..2699703.16 rows=1 width=8)
-> Nested Loop (cost=0.00..2614087.74 rows=34246165 width=0)
Join Filter: (((t1.d * t1.d) + t1.d) > (t2.d + (t2.d * t2.d)))
-> Seq Scan on tt t1 (cost=0.00..22800.00 rows=10136 width=8)
Filter: (ii < 10000)
-> Materialize (cost=0.00..22850.68 rows=10136 width=8)
-> Seq Scan on tt t2 (cost=0.00..22800.00 rows=10136 width=8)
Filter: (ii < 10000)
(8 rows)
Time: 0.561 ms
postgres=# select count(*) from tt t1, tt t2 where t1.ii < 10000 and t2.ii < 10000 and t1.d * t1.d + t1.d > t2.d + t2.d * t2.d;
count
----------
49985001
(1 row)
Time: 19706.890 ms
postgres=# select count(*) from tt t1, tt t2 where t1.ii < 10000 and t2.ii < 10000 and t1.n * t1.n + t1.n > t2.n + t2.n * t2.n;
count
----------
49985001
(1 row)
Time: 70787.289 ms
postgres=# select count(*) from tt t1, tt t2 where t1.ii < 10000 and t2.ii < 10000 and t1.d64 * t1.d64 + t1.d64 > t2.d64 + t2.d64 * t2.d64;
count
----------
49985001
(1 row)
Time: 49861.689 ms
postgres=# select count(*) from tt t1, tt t2 where t1.ii < 10000 and t2.ii < 10000 and t1.d128 * t1.d128 + t1.d128 > t2.d128 + t2.d128 * t2.d128;
count
----------
49985001
(1 row)
Time: 65779.153 ms
小结
PostgreSQL内置的numeric类型属于"无限"精度数值类型，其他类型都是定长存储，使用时不需要调用palloc，效率较高。
如果你要使用超过双精能表示的有效范围的数值，目前只能选择decimal\numeric类型，而这个类型前面说了，由于是变长设计，需要调用palloc，效率一般。
那么在数据分析领域，或者需要处理非常多的数据记录时，numeric类型的开销是较大的。
从前面的测试数据，可以观察到性能最好的是float8，其次是decimal64, decimal64不需要使用palloc，性能比numeric好1.5倍左右，而decimal128也比numeric性能好不少。
期待将来PostgreSQL内置decimal64, decimal128。
参考
《PostgreSQL fixeddecimal - 用CPU "硬解码" 提升1倍 数值运算能力 助力金融大数据量计算》
https://www.postgresql.org/message-id/flat/CAFWGqnsuyOKdOwsNLVtDU1LLjS%3D66xmxxxS8Chnng_zSB5_uCg%40mail.gmail.com#CAFWGqnsuyOKdOwsNLVtDU1LLjS=66xmxxxS8Chnng_zSB5_uCg@mail.gmail.com
https://github.com/vitesse-ftian/pgdecimal
https://pgxn.org/dist/pgdecimal/1.0.0/
https://github.com/2ndQuadrant/fixeddecimal