作者
digoal
日期
2016-03-20
标签
PostgreSQL , BRIN , 流式数据 , 性能优化 , 数据写入 , 批量写入 , 异步提交 , 合理CPU利用 , XLOG锁
背景
本文主要介绍并测试PostgreSQL 在中高端x86服务器上的数据插入速度（目标表包含一个时间字段的索引），帮助企业用户了解PostgreSQL在这类场景下的性能表现。
这类场景常见于 :
运营商网关数据，金融行业数据，产生量大，并且要求快速插入大数据库中持久化保存。
另外, 用户如果需要流式实时处理，可以参考基于PostgreSQL的流式处理方案，一天处理1万亿的实时流式处理是如何实现的？
《"物联网"流式处理应用 - 用PostgreSQL实时处理(万亿每天)》
TEST CASE
1. 平均每条记录长度360字节， (比较常见的长度)
2. 时间字段创建索引。
3. 每轮测试插入12TB数据，插入完12T后清除数据继续插入。循环。
实际场景中，我们的单机很少有PB级的空间，一天写100TB，一个月就是3PB了。所以我们可以借助阿里云的OSS EXT插件，与海量对象存储来存历史数据。常见的架构如下：
用法参考
https://help.aliyun.com/document_detail/44461.html
https://help.aliyun.com/document_detail/35457.html
4. 测试满24小时停止测试。
5. 统计24小时插入的记录数。
TEST 结果
24小时一共完成12轮测试，平均每轮测试耗时7071秒。
506万行/s，1.78 GB/s，全天插入4372亿，154TB数据。
测试的硬件环境
1\. X86服务器
2\. 3?核。
3\. 5??G 内存
4\. 几块SSD，15TB容量
软件环境
1\. CENTOS 6.x x64
2\ .xfs
3\. PostgreSQL 9.5
系统配置参考
https://github.com/digoal/pgsql_admin_script/blob/master/pgsql_perf_tuning.md
数据库配置
详见：
./configure --prefix=/home/digoal/pgsql9.5.1 --with-blocksize=32 --with-segsize=128 --with-wal-blocksize=32 --with-wal-segsize=64
make && make install
PostgreSQL支持hugepage的方法请参考：
https://yq.aliyun.com/articles/8482
参数
listen_addresses = '0.0.0.0' # what IP address(es) to listen on;
fsync=on
port = 1921 # (change requires restart)
max_connections = 600 # (change requires restart)
superuser_reserved_connections = 13 # (change requires restart)
unix_socket_directories = '.' # comma-separated list of directories
unix_socket_permissions = 0700 # begin with 0 to use octal notation
tcp_keepalives_idle = 60 # TCP_KEEPIDLE, in seconds;
tcp_keepalives_interval = 10 # TCP_KEEPINTVL, in seconds;
tcp_keepalives_count = 10 # TCP_KEEPCNT;
shared_buffers = 256GB # min 128kB
huge_pages = on # on, off, or try
work_mem = 512MB # min 64kB
maintenance_work_mem = 1GB # min 1MB
autovacuum_work_mem = 1GB # min 1MB, or -1 to use maintenance_work_mem
dynamic_shared_memory_type = posix # the default is the first option
bgwriter_delay = 10ms # 10-10000ms between rounds
bgwriter_lru_maxpages = 1000 # 0-1000 max buffers written/round
bgwriter_lru_multiplier = 2.0
synchronous_commit = off # synchronization level;
full_page_writes = on # recover from partial page writes
wal_buffers = 2047MB # min 32kB, -1 sets based on shared_buffers
wal_writer_delay = 10ms # 1-10000 milliseconds
checkpoint_timeout = 55min # range 30s-1h
max_wal_size = 512GB
checkpoint_completion_target = 0.9 # checkpoint target duration, 0.0 - 1.0
effective_cache_size = 40GB
log_destination = 'csvlog' # Valid values are combinations of
logging_collector = on # Enable capturing of stderr and csvlog
log_directory = 'pg_log' # directory where log files are written,
log_filename = 'postgresql-%Y-%m-%d_%H%M%S.log' # log file name pattern,
log_file_mode = 0600 # creation mode for log files,
log_truncate_on_rotation = on # If on, an existing log file with the
log_checkpoints = off
log_connections = off
log_disconnections = off
log_error_verbosity = verbose # terse, default, or verbose messages
log_timezone = 'PRC'
log_autovacuum_min_duration = 0 # -1 disables, 0 logs all actions and
datestyle = 'iso, mdy'
timezone = 'PRC'
lc_messages = 'C' # locale for system error message
lc_monetary = 'C' # locale for monetary formatting
lc_numeric = 'C' # locale for number formatting
lc_time = 'C' # locale for time formatting
default_text_search_config = 'pg_catalog.english'
autovacuum=off
创建测试表 :
每32K的block存储89条记录, 每条记录360字节。
```
postgres=# select string_agg(i,'') from (select md5(random()::text) i from generate_series(1,10) t(i)) t(i);
string_agg
53d3ec7adbeacc912a45bdd8557b435be848e4b1050dc0f5e46b75703d4745833541b5dabc177db460b6b1493961fc72c478daaaac74bcc89aec4f946a496028d9cff1cc4144f738e01ea36436455c216aa697d87fe1f87ceb49134a687dc69cba34c9951d0c9ce9ca82bba229d56874af40498dca5f
d8dfb9c877546db76c35a3362d6bdba6472d3919289b6eaeeab58feb4f6e79592fc1dd8253fd4c588a29
(1 row)
postgres=# create unlogged table test(crt_time timestamp, info text default '53d3ec7adbeacc912a45bdd8557b435be848e4b1050dc0f5e46b75703d4745833541b5dabc177db460b6b1493961fc72c478daaaac74bcc89aec4f946a496028d9cff1cc4144f738e01ea36436455c216aa697d87fe1f87ceb49134a687dc69cba34c9951d0c9ce9ca82bba229d56874af40498dca5f
d8dfb9c877546db76c35a3362d6bdba6472d3919289b6eaeeab58feb4f6e79592fc1dd8253fd4c588a29');
postgres=# alter table test alter column info set storage plain;
postgres=# insert into test select now() from generate_series(1,1000);
postgres=# select ctid from test limit 1000;
```
分别在3个物理块设备上创建3个表空间目录，同时在数据库中创建表空间。
tbs1, tbs2, tbs3.
创建多个分表，用于减少 block extend exclusive lock 冲突。
do language plpgsql $$
declare
i int;
sql text;
begin
for i in 1..42 loop
sql := 'create unlogged table test'||i||' (like test including all) tablespace tbs1';
execute sql;
sql := 'create index idx_test'||i||' on test'||i||' using brin (crt_time) with (pages_per_range=512) tablespace tbs1';
execute sql;
end loop;
for i in 43..84 loop
sql := 'create unlogged table test'||i||' (like test including all) tablespace tbs2';
execute sql;
sql := 'create index idx_test'||i||' on test'||i||' using brin (crt_time) with (pages_per_range=512) tablespace tbs2';
execute sql;
end loop;
for i in 85..128 loop
sql := 'create unlogged table test'||i||' (like test including all) tablespace tbs3';
execute sql;
sql := 'create index idx_test'||i||' on test'||i||' using brin (crt_time) with (pages_per_range=512) tablespace tbs3';
execute sql;
end loop;
end;
$$;
又见黑科技 BRIN 索引方法
这里使用的是brin范围索引，PostgreSQL 针对物联网流式数据的黑科技。
postgres=# \di
List of relations
Schema | Name | Type | Owner | Table
--------+-------------+-------+----------+---------
public | idx_test1 | index | postgres | test1
public | idx_test10 | index | postgres | test10
public | idx_test100 | index | postgres | test100
public | idx_test101 | index | postgres | test101
public | idx_test102 | index | postgres | test102
public | idx_test103 | index | postgres | test103
public | idx_test104 | index | postgres | test104
public | idx_test105 | index | postgres | test105
public | idx_test106 | index | postgres | test106
......
......
public | idx_test90 | index | postgres | test90
public | idx_test91 | index | postgres | test91
public | idx_test92 | index | postgres | test92
public | idx_test93 | index | postgres | test93
public | idx_test94 | index | postgres | test94
public | idx_test95 | index | postgres | test95
public | idx_test96 | index | postgres | test96
public | idx_test97 | index | postgres | test97
public | idx_test98 | index | postgres | test98
public | idx_test99 | index | postgres | test99
(128 rows)
生成测试脚本, 一个连接一次插入178条记录，占用2个32KB的block :
```
vi test.sql
insert into test(crt_time) values (now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now()),(now());
for ((i=1;i<=128;i++)) do sed "s/test/test$i/" test.sql > ./test$i.sql; done
```
开始测试前清除数据：
do language plpgsql $$
declare
i int;
sql text;
begin
for i in 1..128 loop
sql := 'truncate test'||i;
execute sql;
end loop;
end;
$$;
测试方法：
每轮测试插入12TB数据。通过以下方式控制：
1. 使用128个并行连接，每个连接执行1572864个事务。
2. 一共执行201326592个事务（每个事务插入178条记录）。
3. 一共插入35836133376条记录（358.36 亿记录）（共计12TB 数据，索引空间另算）。
进行下一轮测试前，输出日志，并TRUNCATE所有的数据，然后重复以上测试。
直到测试满24小时，输出统计数据。
测试脚本如下 :
```
vi test.sh
!/bin/bash
if [ $# -ne 5 ]; then
echo "please use: $0 ip port dbname user pwd"
exit 1
fi
IP=$1
PORT=$2
DBNAME=$3
USER=$4
PASSWORD=$5
export PGPASSWORD=$PASSWORD
DEP_CMD="psql"
which $DEP_CMD
if [ $? -ne 0 ]; then
echo -e "dep commands: $DEP_CMD not exist."
exit 1
fi
truncate() {
psql -h $IP -p $PORT -U $USER $DBNAME <<EOF
do language plpgsql \$\$
declare
i int;
sql text;
begin
for i in 1..128 loop
sql := 'truncate test'||i;
execute sql;
end loop;
end;
\$\$;
checkpoint;
\q
EOF
}
truncate data first
truncate
START=date +%s
echo "date +%F%T $START"
for ((x=1;x>0;x++))
do
------------------------------------------------------
echo "Round $x test start: date +%F%T date +%s"
for ((i=1;i<=128;i++))
do
pgbench -M prepared -n -r -f ./test$i.sql -h $IP -p $PORT -U $USER $DBNAME -c 1 -j 1 -t 1572864 >>./$i.log 2>&1 &
done
wait
echo "Round $x test end: date +%F%T date +%s"
------------------------------------------------------
if [ $((date +%s-$START)) -gt 86400 ]; then
echo "end date +%F%T date +%s"
echo "duration second: $((date +%s-$START))"
exit 0
fi
echo "Round $x test end, start truncate date +%F%T date +%s"
truncate
echo "Round $x test end, end truncate date +%F%T date +%s"
done
```
测试
nohup ./test.sh xxx.xxx.xxx.xxx 1921 postgres postgres postgres >./test.log 2>&1 &
测试结果
24小时完成12轮测试，平均每轮测试耗时7071秒。
506万行/s（每行360字节），1.78GB/s，全天插入4372亿，154TB数据。
查询性能
```
postgres=# select min(crt_time),max(crt_time) from test1;
min | max
----------------------------+----------------------------
2016-04-08 00:32:26.842728 | 2016-04-08 02:29:41.583367
(1 row)
postgres=# explain select count(*) from test1 where crt_time between '2016-04-08 00:32:00' and '2016-04-08 00:33:00';
QUERY PLAN
Aggregate (cost=1183919.81..1183919.82 rows=1 width=0)
-> Bitmap Heap Scan on test1 (cost=14351.45..1180420.19 rows=1399849 width=0)
Recheck Cond: ((crt_time >= '2016-04-08 00:32:00'::timestamp without time zone) AND (crt_time <= '2016-04-08 00:33:00'::timestamp without time zone))
-> Bitmap Index Scan on idx_test1 (cost=0.00..14001.49 rows=1399849 width=0)
Index Cond: ((crt_time >= '2016-04-08 00:32:00'::timestamp without time zone) AND (crt_time <= '2016-04-08 00:33:00'::timestamp without time zone))
(5 rows)
Time: 0.382 ms
postgres=# select count(*) from test1 where crt_time between '2016-04-08 00:32:00' and '2016-04-08 00:33:00';
count
2857968
(1 row)
Time: 554.474 ms
```
如何潇洒地做到每天写入百TB数据
1、合理的环境部署和数据库部署，请参考
2、异步提交
set synchronous_commit=off;
3、关闭被写入表的autovacuum_enable和toast.autovacuum_enable开关。建议不需要进行海量输入的表，不要关闭autovacuum。
postgres=# alter table test_1 set (autovacuum_enabled =off);
ALTER TABLE
postgres=# alter table test_1 set (toast.autovacuum_enabled =off);
ALTER TABLE
4、使用UNLOGGED TABLE（不建议生产使用）
避免XLOG writer lock瓶颈。
5、使用2-4倍CPU核数的连接
例如32核，建议64-128个连接。少了发挥不了CPU的性能，多了无益（反而增加CPU时间片切换带来的开销）。
6、每个连接写不同的表。
减少表的extend file exclusive lock的锁冲突。如果不好实现的话，可以使用PostgreSQL提供的动态函数功能。在《PostgreSQL on ECS多云盘的部署、快照备份和恢复》有介绍。
7、使用批量提交（COPY或INSERT INTO TABLE VALUES (),(),...();）
批量提交减少交互开销，减少代码路径，提高性能。
8、如果需要使用时序查询，建议使用BRIN索引，不要使用B-TREE索引。请参考
《PostgreSQL 物联网黑科技 - 瘦身几百倍的索引(BRIN index)》
9、如果你的业务需要实时的分析。参考《"物联网"流式处理应用 - 用PostgreSQL实时处理(万亿每天)》
另一组测试数据，以及对应的测试方法参考：
《PostgreSQL on ECS多云盘的部署、快照备份和恢复》
将PostgreSQL部署在ECS虚拟机上，也已经接近每秒千万行、一天接近1万亿行的写入速度。
里面附录了数据结构的定义、批量写入的方法、动态函数等。
小结
1. 这个CASE主要的应用场景是实时的大数据入库，例如 物联网 的应用场景，大量的 传感器 会产生庞大的数据。
又比如传统的 运营商网关 ，也会有非常庞大的流量数据或业务数据需要实时的入库。
索引方面，用到了PostgreSQL黑科技BRIN。
2. 除了实时入库，用户如果需要流式实时处理，可以参考基于PostgreSQL的流式处理方案，
一天处理1万亿的实时流式处理是如何实现的？
《"物联网"流式处理应用 - 用PostgreSQL实时处理(万亿每天)》
3. 瓶颈, 还是在IO上面 , 有几个表现，TOP大量进程处于D(front io)状态 。
w: S -- Process Status
The status of the task which can be one of:
’D’ = uninterruptible sleep
’R’ = running
’S’ = sleeping
’T’ = traced or stopped
’Z’ = zombie
所有块设备的使用率均达100% 。
清理数据时 ：
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
dfa 0.00 0.00 5807.39 167576.65 1464080.93 1340613.23 16.18 535.69 3.02 0.01 116.77
dfb 0.00 0.00 5975.10 185132.68 1506714.40 1481061.48 15.63 459.46 2.32 0.01 110.62
dfc 0.00 0.00 5715.56 182584.05 1440771.98 1460672.37 15.41 568.02 2.93 0.01 112.37
插入数据时 ：
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
dfa 0.00 0.00 0.00 235936.00 0.00 1887488.00 8.00 2676.34 11.17 0.00 99.10
dfb 0.00 0.00 0.00 237621.00 0.00 1900968.00 8.00 66.02 0.10 0.00 99.10
dfc 0.00 0.00 0.00 239830.00 0.00 1918632.00 8.00 10.66 0.04 0.00 101.30
IO层面的性能问题，可以通过优化代码（例如 PostgreSQL bgwriter 在写出数据时，尽量顺序写出），便于OS层进行IO合并，来缓解IO压力，从这个信息来看，单次写IO的大小还可以再大点。
有几个工具你可能用得上，perf, systemtap, goprof.
如果要较全面的分析，建议把PostgreSQL --enable-profiling打开用于诊断。
PostgreSQL 许愿链接
您的愿望将传达给PG kernel hacker、数据库厂商等, 帮助提高数据库产品质量和功能, 说不定下一个PG版本就有您提出的功能点. 针对非常好的提议，奖励限量版PG文化衫、纪念品、贴纸、PG热门书籍等，奖品丰富，快来许愿。开不开森.
9.9元购买3个月阿里云RDS PostgreSQL实例
PostgreSQL 解决方案集合
德哥 / digoal's github - 公益是一辈子的事.
