本文共 12093 字,大约阅读时间需要 40 分钟。
关于PostgreSQL的性能调优可以参考《PostgreSQL 9.0 High Performance》,以及朱贤文在2014 PostgreSQL中国用户大会上分享的《高性能Postgres 最佳实践》。当然,首先还是应该看看PostgreSQL手册的相关章节。我们在调优时不必每个细节都做到最优,抓住主要矛盾即可。因为有些东西不在你的控制之下,或者那样优化之后维护起来麻烦。下面尝试在虚机下进行快速的PostgreSQL参数调优。
CPU: 4 core
Mem: 8G OS: CentOS 6.3(64 Bit)PostgreSQL:9.4.5
sysbench:0.4.12
文件系统:ext4
基本的性能参数设置可以利用下面这个在线小工具评估。
填入系统信息,并固定最大连接数为300后,选择不同DB Type,这个工具会给出不同的参数。
Web applications
max_connections = 300shared_buffers = 2GBeffective_cache_size = 6GBwork_mem = 6990kBmaintenance_work_mem = 512MBcheckpoint_segments = 32checkpoint_completion_target = 0.7wal_buffers = 16MBdefault_statistics_target = 100
Online transaction processing systems
max_connections = 300shared_buffers = 2GBeffective_cache_size = 6GBwork_mem = 6990kBmaintenance_work_mem = 512MBcheckpoint_segments = 64checkpoint_completion_target = 0.9wal_buffers = 16MBdefault_statistics_target = 100
Data warehouses
max_connections = 300shared_buffers = 2GBeffective_cache_size = 6GBwork_mem = 3495kBmaintenance_work_mem = 1GBcheckpoint_segments = 128checkpoint_completion_target = 0.9wal_buffers = 16MBdefault_statistics_target = 500
上面3种DB Type,越往后写越重,checkpoint的频率也调得越低。由于后面要做OLTP的性能评估,所以选用Online transaction processing systems的设置。
综合考虑log等需求,初步在postgresql.conf中设置参数如下
listen_addresses = '*'port = 5432max_connections = 300shared_buffers = 2GBeffective_cache_size = 6GBwork_mem = 6990kBmaintenance_work_mem = 512MBcheckpoint_segments = 64checkpoint_completion_target = 0.9wal_buffers = 16MBdefault_statistics_target = 100logging_collector = onlog_directory = 'pg_log'log_filename = 'postgresql-%Y-%m-%d_%H%M%S.log'log_truncate_on_rotation = onlog_rotation_age = 1440log_rotation_size = 100000log_line_prefix='%m %p %x'wal_level = hot_standby
流复制时需要设置wal_level = hot_standby,单机场景下可以设置其它值以输出更少的WAL日志。
下面这几个参数有些需要在性能和持久性之间平衡,先全部采用默认值。
wal_sync_method = fsynccommit_delay = 0synchronous_commit = onfull_page_writes = onfsync = on
用sysbench 做oltp的性能测试,不管使用simple还是complex测试模式,sysbench prepare时创建单个相同的测试表(那就不可能测到join了)。
建表语句如下
CREATE TABLE sbtest (id SERIAL NOT NULL , k integer DEFAULT '0' NOT NULL, c char(120) DEFAULT '' NOT NULL, pad char(60) DEFAULT '' NOT NULL, PRIMARY KEY (id) ) CREATE INDEX k on sbtest(k) |
然后插入数据,插入数据量由oltp-table-size决定。
INSERT INTO sbtest(k, c, pad) VALUES (0,' ','qqqqqqqqqqwwwwwwwwwweeeeeeeeeerrrrrrrrrrtttttttttt') |
当插入数据量很大时,由于每个插入语句都是一次提交,因此速度很慢。我们可以用下面的方法手动准备数据,而不是使用sysbench的prepare。
DROP TABLE sbtest; CREATE TABLE sbtest (id SERIAL NOT NULL , k integer DEFAULT '0' NOT NULL, c char(120) DEFAULT '' NOT NULL, pad char(60) DEFAULT '' NOT NULL, PRIMARY KEY (id) ); CREATE INDEX k on sbtest(k); INSERT INTO sbtest(k, c, pad) select 0,' ','qqqqqqqqqqwwwwwwwwwweeeeeeeeeerrrrrrrrrrtttttttttt' from generate_series(1,5000000); |
插入500万记录后,数据表大小大约为1GB。
postgres=# \d+ List of relations Schema | Name | Type | Owner | Size | Description --------+------------------+----------+----------+------------+------------- public | sbtest | table | postgres | 1056 MB | public | sbtest_id_seq | sequence | postgres | 8192 bytes | (2 rows) |
simple模式的测试语句如下
SELECT c from sbtest where id=$1 |
$1是个取值在oltp-table-size范围内的随机数,随机数的生成算法由oltp-dist-type决定,包括uniform,gaussian,special三种,默认是special,生成的随机数有75%(由oltp-dist-pct控制)集中在一个1%(由oltp-dist-pct控制)的热点区域。
complex模式且oltp-read-only=on时的测试语句如下
BEGIN SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id between $1 and $2 SELECT SUM(K) from sbtest where id between $1 and $2 SELECT c from sbtest where id between $1 and $2 order by c SELECT DISTINCT c from sbtest where id between $1 and $2 order by c COMMIT |
上面的between范围查询,范围大小为100(由oltp-range-size控制)。
complex模式且oltp-read-only=off时的测试语句如下
BEGIN SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id=$1 SELECT c from sbtest where id between $1 and $2 SELECT SUM(K) from sbtest where id between $1 and $2 SELECT c from sbtest where id between $1 and $2 order by c SELECT DISTINCT c from sbtest where id between $1 and $2 order by c UPDATE sbtest set k=k+1 where id=$1 UPDATE sbtest set c=$1 where id=$2 UPDATE sbtest set k=k+1 where id=$1 DELETE from sbtest where id=$1 INSERT INTO sbtest values($1,0,' ','aaaaaaaaaaffffffffffrrrrrrrrrreeeeeeeeeeyyyyyyyyyy') COMMIT |
但是在并发数很高的情况下,会报下面的错误
ERROR: duplicate key value violates unique constraint "sbtest_pkey"
为了回避这个问题,参考德哥的方法临时修改sysbench的代码(这个问题和PostgreSQL的MVCC实现机制有关,详见)。
修改方法如下:
修改sysbench-0.4.12/sysbench/tests/oltp/sb_oltp.c
找到
/* Prepare the insert statement */ snprintf(query, MAX_QUERY_LEN, "INSERT INTO %s values(?,0,' '," "'aaaaaaaaaaffffffffffrrrrrrrrrreeeeeeeeeeyyyyyyyyyy')", args.table_name); |
改成
/* Prepare the insert statement */ if (args.auto_inc) snprintf(query, MAX_QUERY_LEN, "INSERT INTO %s(k,c,pad) values(0,' '," "'aaaaaaaaaaffffffffffrrrrrrrrrreeeeeeeeeeyyyyyyyyyy')", args.table_name); else snprintf(query, MAX_QUERY_LEN, "INSERT INTO %s values(?,0,' '," "'aaaaaaaaaaffffffffffrrrrrrrrrreeeeeeeeeeyyyyyyyyyy')", args.table_name); |
参考:
修改后,最后一条INSERT句就变成了下面这样了
INSERT INTO sbtest(k,c,pad) values(0,' ','aaaaaaaaaaffffffffffrrrrrrrrrreeeeeeeeeeyyyyyyyyyy') |
注:其实更加正确的做法应该是使用sysbench 0.5而不是0.4,sysbench 0.5没有这个问题,而且0.5支持lua脚本,支持的测试方式更灵活。
simple模式
执行下面的命令,100并发,TPS为29200.43,测试时CPU被占满,达到CPU极限。
[postgres@node1 ~]$ sysbench --test=oltp --db-driver=pgsql --pgsql-host=127.0.0.1 --pgsql-port=5432 --pgsql-user=postgres --pgsql-password=postgres --pgsql-db=postgres --oltp-table-size=5000000 --num-threads=100 --max-requests=0 --max-time=60 --oltp-test-mode=simple --oltp-read-only=on runsysbench 0.4.12: multi-threaded system evaluation benchmarkRunning the test with following options:Number of threads: 100Doing OLTP test.Running simple OLTP testDoing read-only testUsing Special distribution (12 iterations, 1 pct of values are returned in 75 pct cases)Using "BEGIN" for starting transactionsUsing auto_inc on the id columnThreads started!Time limit exceeded, exiting...(last message repeated 99 times)Done.OLTP test statistics: queries performed: read: 1752165 write: 0 other: 0 total: 1752165 transactions: 1752165 (29200.43 per sec.) deadlocks: 0 (0.00 per sec.) read/write requests: 1752165 (29200.43 per sec.) other operations: 0 (0.00 per sec.)Test execution summary: total time: 60.0048s total number of events: 1752165 total time taken by event execution: 5993.5646 per-request statistics: min: 0.04ms avg: 3.42ms max: 968.42ms approx. 95 percentile: 0.34msThreads fairness: events (avg/stddev): 17521.6500/2549.74 execution time (avg/stddev): 59.9356/0.01
complex模式且oltp-read-only=on
执行下面的命令,100并发,TPS为1505.38,测试时CPU被占满,达到CPU极限,其中sysbench进程的CPU率大约80%。
[postgres@node1 ~]$ sysbench --test=oltp --db-driver=pgsql --pgsql-host=127.0.0.1 --pgsql-port=5432 --pgsql-user=postgres --pgsql-password=postgres --pgsql-db=postgres --oltp-table-size=5000000 --num-threads=100 --max-requests=0 --max-time=60 --oltp-test-mode=complex --oltp-read-only=on runsysbench 0.4.12: multi-threaded system evaluation benchmarkRunning the test with following options:Number of threads: 100Doing OLTP test.Running mixed OLTP testDoing read-only testUsing Special distribution (12 iterations, 1 pct of values are returned in 75 pct cases)Using "BEGIN" for starting transactionsUsing auto_inc on the id columnThreads started!Time limit exceeded, exiting...(last message repeated 99 times)Done.OLTP test statistics: queries performed: read: 1265264 write: 0 other: 180752 total: 1446016 transactions: 90376 (1505.38 per sec.) deadlocks: 0 (0.00 per sec.) read/write requests: 1265264 (21075.27 per sec.) other operations: 180752 (3010.75 per sec.)Test execution summary: total time: 60.0355s total number of events: 90376 total time taken by event execution: 5998.5193 per-request statistics: min: 1.98ms avg: 66.37ms max: 4032.00ms approx. 95 percentile: 161.49msThreads fairness: events (avg/stddev): 903.7600/74.26 execution time (avg/stddev): 59.9852/0.04
complex模式且oltp-read-only=off
执行下面的命令,100并发,TPS为1232.77,测试时CPU被占满,其中sysbench进程的CPU率大约80%,磁盘Busy大约10+%,磁盘写入也是10+MB/s。
[postgres@node1 ~]$ sysbench --test=oltp --db-driver=pgsql --pgsql-host=127.0.0.1 --pgsql-port=5432 --pgsql-user=postgres --pgsql-password=postgres --pgsql-db=postgres --oltp-table-size=5000000 --num-threads=100 --max-requests=0 --max-time=60 --oltp-test-mode=complex --oltp-read-only=off runsysbench 0.4.12: multi-threaded system evaluation benchmarkRunning the test with following options:Number of threads: 100Doing OLTP test.Running mixed OLTP testUsing Special distribution (12 iterations, 1 pct of values are returned in 75 pct cases)Using "BEGIN" for starting transactionsUsing auto_inc on the id columnThreads started!Time limit exceeded, exiting...(last message repeated 99 times)Done.OLTP test statistics: queries performed: read: 1035986 write: 369995 other: 147998 total: 1553979 transactions: 73999 (1232.77 per sec.) deadlocks: 0 (0.00 per sec.) read/write requests: 1405981 (23422.71 per sec.) other operations: 147998 (2465.55 per sec.)Test execution summary: total time: 60.0264s total number of events: 73999 total time taken by event execution: 5998.7493 per-request statistics: min: 2.69ms avg: 81.07ms max: 547.21ms approx. 95 percentile: 199.32msThreads fairness: events (avg/stddev): 739.9900/23.37 execution time (avg/stddev): 59.9875/0.04
由于上面测试场景的性能瓶颈都在CPU,基本没什么可优化的,(实际试了下,确实也没有优化效果)。如果瓶颈在刷盘可以通过调节下面几个参数优化。
No | 参数 | 说明 | 风险 |
---|---|---|---|
1 | wal_sync_method = fdatasync | WAL刷盘的系统调用,根据之前的测试fdatasync比fsync性能好。但是Linux下的默认值就是fdatasync所以也不同修改。 | 无风险 |
2 | commit_delay = 100 | 提交延迟,单位是微妙(不是毫秒),以进行组提交。 | 设得太大会影响事务的响应时间。 注)实际压测发现效果不稳定。因并发连接数,热数据分布,commit_delay等的不同,有时性能提高有时降低,所以优化这个值要根据实际的应用环境。 |
3 | synchronous_commit = off | 异步提交,WAL刷盘交给OS | OS crash时,可能丢失最近的提交。 |
4 | full_page_writes = off | 在WAL中checkpoint后的第一次修改page时不写全page数据 | 对不支持原子写的文件系统或存储设备上,OS crash时,数据文件会损坏。 |
5 | fsync = off | 不刷盘,完全交给OS刷 | OS crash时,数据文件可能会损坏。 |
还有一种经常被提到的优化方法是修改page大小,PostgreSQL默认的page大小是8K,编译时可以指定。
./configure --with-blocksize=16
下面是16K和8K的性能对比,从下面可以看出16K反而性能变差,所以16K的优化方法一定要看机器环境和场景的(可能更适合OLAP吧)。而且改page大小要重新编译源码,8K的数据目录和16K的又不兼容,不能互相复制。所以修改page大小要慎重。
100线程,500W数据的sysbench测试出的tps值:
test mode | 8K | 16K |
---|---|---|
simple | 29200.43 | 28795.46 |
complex模式且oltp-read-only=on | 1505.38 | 1462.40 |
complex模式且oltp-read-only=off | 1232.77 | 1058.48 |
转载地址:http://nilox.baihongyu.com/