Postgresql优化案例

环境

参考德哥文章和自己的一些实践总结

原文：https://github.com/digoal/blog/blob/master/201203/20120313_01.md

硬件

CPU: Intel(R) Xeon(R) CPU E5-2682 v4 @ 2.50GHz - 64cores
SSD: INTEL SSDSCKHB34
HDD: ST8000NM0055-1RM
MEM: 516754MB

软件

postgres10.1
pgbench

全部使用域套接字连接，避免网络影响

基准测试[0]

测试模型

数据库默认参数模版

create database etest;
\c etest

CREATE TABLE user_info (userid int, engname text, cnname text, occupation text, birthday date, signname text, email text, qq numeric, crt_time TIMESTAMP WITHOUT time ZONE, mod_time TIMESTAMP WITHOUT time ZONE);

CREATE TABLE user_session (userid int, logintime TIMESTAMP(0) WITHOUT time ZONE, login_count bigint DEFAULT 0, logouttime TIMESTAMP(0) WITHOUT time ZONE, online_interval interval DEFAULT interval '0');

CREATE TABLE user_login_rec (userid int, login_time TIMESTAMP WITHOUT time ZONE, ip inet);

CREATE TABLE user_logout_rec (userid int, logout_time TIMESTAMP WITHOUT time ZONE, ip inet);

初始化数据

insert into user_info (userid,engname,cnname,occupation,birthday,signname,email,qq,crt_time,mod_time)  
select generate_series(1,20000000),  'digoal.zhou',  '德哥',  'DBA',  '1970-01-01'  ,E'公益是一辈子的事, I\'m Digoal.Zhou, Just do it!',  '[email protected]',  276732431,  clock_timestamp(),  NULL;  
  
insert into user_session (userid) select generate_series(1,20000000);  
  
set work_mem='2048MB';  
set maintenance_work_mem='2048MB';  
alter table user_info add constraint pk_user_info primary key (userid);  
alter table user_session add constraint pk_user_session primary key (userid);  

模拟业务场景登陆、登出

create or replace function f_user_login   
(i_userid int,  
OUT o_userid int,  
OUT o_engname text,  
OUT o_cnname text,  
OUT o_occupation text,  
OUT o_birthday date,  
OUT o_signname text,  
OUT o_email text,  
OUT o_qq numeric  
)  
as $BODY$  
declare  
begin  
select userid,engname,cnname,occupation,birthday,signname,email,qq  
into o_userid,o_engname,o_cnname,o_occupation,o_birthday,o_signname,o_email,o_qq  
from user_info where userid=i_userid;  
insert into user_login_rec (userid,login_time,ip) values (i_userid,now(),inet_client_addr());  
update user_session set logintime=now(),login_count=login_count+1 where userid=i_userid;  
return;  
end;  
$BODY$  
language plpgsql;  

create or replace function f_user_logout  
(i_userid int,  
OUT o_result int  
)  
as $BODY$  
declare  
begin  
insert into user_logout_rec (userid,logout_time,ip) values (i_userid,now(),inet_client_addr());  
update user_session set logouttime=now(),online_interval=online_interval+(now()-logintime) where userid=i_userid;  
o_result := 0;  
return;  
exception   
when others then  
o_result := 1;  
return;  
end;  
$BODY$  
language plpgsql;  

pgbench测试脚本login

\set userid random(1,20000000)
select userid,engname,cnname,occupation,birthday,signname,email,qq from user_info where userid=:userid;
insert into user_login_rec (userid,login_time,ip) values (:userid,now(),inet_client_addr());
update user_session set logintime=now(),login_count=login_count+1 where userid=:userid;

pgbench测试脚本logout

\set userid random(1,20000000)
insert into user_logout_rec (userid,logout_time,ip) values (:userid,now(),inet_client_addr());
update user_session set logouttime=now(),online_interval=online_interval+(now()-logintime) where userid=:userid;

测试command

# client 8 threads 8 time 180s vacuum before test
pgbench -M simple -r -c 8 -f ./login.sql -j 8 -n -T 180 -P 1 -v etest > /tmp/pb_login_j8_0.log 2>&1

测试结果:312

transaction type: ./login.sql
scaling factor: 1
query mode: simple
number of clients: 8
number of threads: 8
duration: 180 s
number of transactions actually processed: 56364
latency average = 25.582 ms
latency stddev = 56.738 ms
tps = 312.545965 (including connections establishing)
tps = 312.550425 (excluding connections establishing)
script statistics:
 - statement latencies in milliseconds:
         0.002  \set userid random(1,20000000)
         0.209  select userid,engname,cnname,occupation,birthday,signname,email,qq from user_info where userid=:userid;
        12.635  insert into user_login_rec (userid,login_time,ip) values (:userid,now(),inet_client_addr());
        12.737  update user_session set logintime=now(),login_count=login_count+1 where userid=:userid;

优化阶段[1] - 开启异步提交

瓶颈分析

测试发现IO瓶颈比较明显

iostat sdb -x 1 10

avg-cpu:  %user   %nice %system %iowait  %steal   %idle
           0.16    0.00    0.11    0.73    0.00   99.00

Device:         rrqm/s   wrqm/s     r/s     w/s   rsec/s   wsec/s avgrq-sz avgqu-sz   await  svctm  %util
sdb               0.00   873.00    0.00  678.00     0.00 13256.00    19.55    48.62   87.26   1.47  99.60

%user：CPU处在用户模式下的时间百分比。
%nice：CPU处在带NICE值的用户模式下的时间百分比。
%system：CPU处在系统模式下的时间百分比。
%iowait：CPU等待输入输出完成时间的百分比。
%steal：管理程序维护另一个虚拟处理器时，虚拟CPU的无意识等待时间百分比。
%idle：CPU空闲时间百分比。
rrqm/s:  每秒进行 merge 的读操作数目。即 rmerge/s
wrqm/s:  每秒进行 merge 的写操作数目。即 wmerge/s
r/s:  每秒完成的读 I/O 设备次数。即 rio/s
w/s:  每秒完成的写 I/O 设备次数。即 wio/s
rsec/s:  每秒读扇区数。即 rsect/s
wsec/s:  每秒写扇区数。即 wsect/s
rkB/s:  每秒读K字节数。是 rsect/s 的一半，因为每扇区大小为512字节。
wkB/s:  每秒写K字节数。是 wsect/s 的一半。
avgrq-sz:  平均每次设备I/O操作的数据大小 (扇区)。
avgqu-sz:  平均I/O队列长度。
await:  平均每次设备I/O操作的等待时间 (毫秒)。
svctm: 平均每次设备I/O操作的服务时间 (毫秒)。
%util:  一秒中有百分之多少的时间用于 I/O 操作，即被io消耗的cpu百分比

备注：如果 %util 接近 100%，说明产生的I/O请求太多，I/O系统已经满负荷，该磁盘可能存在瓶颈。如果 svctm 比较接近 await，说明 I/O 几乎没有等待时间；如果 await 远大于 svctm，说明I/O 队列太长，io响应太慢，则需要进行必要优化。如果avgqu-sz比较大，也表示有当量io在等待。

优化

sed -ir "s/#*synchronous_commit.*/synchronous_commit = off/" /home/mingjie.gmj/databases/data/pgdata8410/postgresql.conf

pg_ctl restart -m fast
psql -c 'show synchronous_commit'
 synchronous_commit
--------------------
 off

测试

pgbench -M simple -r -c 8 -f ./login.sql -j 8 -n -T 180 -P 1 -v etest > /tmp/pb_login_j8_1.log 2>&1

测试结果:717

transaction type: ./login.sql
scaling factor: 1
query mode: simple
number of clients: 8
number of threads: 8
duration: 180 s
number of transactions actually processed: 132335
latency average = 11.151 ms
latency stddev = 267.181 ms
tps = 717.392391 (including connections establishing)
tps = 717.402489 (excluding connections establishing)
script statistics:
 - statement latencies in milliseconds:
         0.002  \set userid random(1,20000000)
         0.224  select userid,engname,cnname,occupation,birthday,signname,email,qq from user_info where userid=:userid;
         5.579  insert into user_login_rec (userid,login_time,ip) values (:userid,now(),inet_client_addr());
         4.950  update user_session set logintime=now(),login_count=login_count+1 where userid=:userid;

对比[1] —> [0]

set grid
set xlabel "collection time"
set xrange[0:180]
set ylabel "qps"
plot "pb_login_j8_0.log" using 2:4 w l lc 1 lw 1 title "[0]" smooth sbezier,\
"pb_login_j8_1.log" using 2:4 w l lc 2 lw 1 title "[1]" smooth sbezier

优化阶段[2] - 使用extended协议

瓶颈分析

• simple协议：一个简单查询周期是由前端发送一条Query消息给后端进行初始化的。这条消息包含一个用
  文本字符串表达的 SQL 命令(或者一些命令)。 后端根据查询命令串的内容发送一条或者
  更多条响应消息给前端，并且最后是一条ReadyForQuery响应消息。ReadyForQuery通知前端
  它可以安全地发送新命令了。
• 扩展查询：扩展查询协议把上面描述的简单协议分裂成若干个步骤。准备步骤的结果可以被多次复用以
  提高效率。另外，还可以获得额外的特性， 比如可以把数据值作为独立的参数提供而不是必
  须把它们直接插入一个查询字符串。

优化

extended协议

测试

pgbench -M extended -r -c 8 -f ./login.sql -j 8 -n -T 180 -P 1 -v etest > /tmp/pb_login_j8_2.log 2>&1

测试结果:846

transaction type: ./login.sql
scaling factor: 1
query mode: extended
number of clients: 8
number of threads: 8
duration: 180 s
number of transactions actually processed: 152398
latency average = 9.452 ms
latency stddev = 241.005 ms
tps = 846.397999 (including connections establishing)
tps = 846.410210 (excluding connections establishing)
script statistics:
 - statement latencies in milliseconds:
         0.001  \set userid random(1,20000000)
         0.226  select userid,engname,cnname,occupation,birthday,signname,email,qq from user_info where userid=:userid;
         4.800  insert into user_login_rec (userid,login_time,ip) values (:userid,now(),inet_client_addr());
         4.135  update user_session set logintime=now(),login_count=login_count+1 where userid=:userid;

对比

set grid
set xlabel "collection time"
set xrange[0:180]
set ylabel "qps"
plot "pb_login_j8_0.log" using 2:4 w l lc 1 lw 1 title "[0]" smooth sbezier,\
"pb_login_j8_1.log" using 2:4 w l lc 2 lw 1 title "[1]" smooth sbezier,\
"pb_login_j8_2.log" using 2:4 w l lc 3 lw 3 title "[2]" smooth sbezier

优化阶段[3] - 使用prepared协议

http://www.postgres.cn/docs/10/protocol-flow.html#PROTOCOL-FLOW-EXT-QUERY

测试

pgbench -M prepared -r -c 8 -f ./login.sql -j 8 -n -T 180 -P 1 -v etest > /tmp/pb_login_j8_3.log 2>&1

测试结果:932

transaction type: ./login.sql
scaling factor: 1
query mode: prepared
number of clients: 8
number of threads: 8
duration: 180 s
number of transactions actually processed: 169299
latency average = 8.581 ms
latency stddev = 232.825 ms
tps = 932.300127 (including connections establishing)
tps = 932.313787 (excluding connections establishing)
script statistics:
 - statement latencies in milliseconds:
         0.001  \set userid random(1,20000000)
         0.132  select userid,engname,cnname,occupation,birthday,signname,email,qq from user_info where userid=:userid;
         4.407  insert into user_login_rec (userid,login_time,ip) values (:userid,now(),inet_client_addr());
         3.754  update user_session set logintime=now(),login_count=login_count+1 where userid=:userid;

对比

set grid
set xlabel "collection time"
set xrange[0:180]
set ylabel "qps"
plot "pb_login_j8_0.log" using 2:4 w l lc 1 lw 1 title "[0]" smooth sbezier,\
"pb_login_j8_1.log" using 2:4 w l lc 2 lw 1 title "[1]" smooth sbezier,\
"pb_login_j8_2.log" using 2:4 w l lc 3 lw 1 title "[2]" smooth sbezier,\
"pb_login_j8_3.log" using 2:4 w l lc 4 lw 3 title "[3]" smooth sbezier

优化阶段[4] - 使用函数

http://www.postgres.cn/docs/10/protocol-flow.html#idp57579680

优化

loginfunc.sql

\set userid random(1,20000000)
SELECT f_user_login(:userid);

测试

pgbench -M prepared -r -c 8 -f ./loginfunc.sql -j 8 -n -T 180 -P 1 -v etest > /tmp/pb_login_j8_4.log 2>&1

测试结果：987

transaction type: ./loginfunc.sql
scaling factor: 1
query mode: prepared
number of clients: 8
number of threads: 8
duration: 180 s
number of transactions actually processed: 182872
latency average = 8.100 ms
latency stddev = 230.179 ms
tps = 987.645074 (including connections establishing)
tps = 987.658690 (excluding connections establishing)
script statistics:
 - statement latencies in milliseconds:
         0.001  \set userid random(1,20000000)
         7.750  SELECT f_user_login(:userid);

对比

set grid
set xlabel "collection time"
set xrange[0:180]
set ylabel "qps"
set yrange[0:5000]
plot "pb_login_j8_0.log" using 2:4 w l lc 1 lw 1 title "[0]" smooth sbezier,\
"pb_login_j8_1.log" using 2:4 w l lc 2 lw 1 title "[1]" smooth sbezier,\
"pb_login_j8_2.log" using 2:4 w l lc 3 lw 1 title "[2]" smooth sbezier,\
"pb_login_j8_3.log" using 2:4 w l lc 4 lw 1 title "[3]" smooth sbezier,\
"pb_login_j8_3.log" using 2:4 w l lc 5 lw 4 title "[4]" smooth sbezier

优化方法总结

这里是一些常规的优化方法，没项都可以展开一篇文章，这里先记录下，后面在总结。

• 使用SSD

• 异步提交

• 增加并发，经验值当活跃的进程数等于核数的2倍时可以发挥CPU的最大能力

• 批次提交

• 关闭pg_log

• 使用prepared协议

• 使用函数封装业务逻辑

• COPY

• 拆表

• hotstandby读写分离

• 分区表（主表并发有性能问题）

• 连接池：短连接大量并发