背景

PostgreSQL 在向和纵向的扩展能力在开源数据库中一直处于非常领先的地位，例如今年推出的9.6，内置了sharding的功能，同时在scale-up的能力也有非常明显的提升，特别是在多核与高并发处理这块。

社区有同学在128核的机器上测试tpc-b的select only模式可以达到几百万的qps，机器的CPU资源被吃光光。

天下大势，分久必合，合久必分。谈了这么多年的sharding，业务也妥协了这么多年（比如不允许跨shard JOIN，忍受分片不平衡的痛楚，必须要有分区键值，分布式事务，分布式事务一致性等限制或使用门槛）。一个数据库能解决的为什么要分片呢？

原来说用分片去大机，去O，初衷是什么？其实还是太贵对吧。

如今X86的性能已经非常好，SSD也非常廉价，给PostgreSQL一台顶级的X86，能把机器的硬件资源掏空，换来的是非常优秀的性能，还有对应用完全自由的使用，不再受shard的多种约束束缚。

除了读的高并发有明显的性能提升，在写这块，引入了动态扩展数据文件，从而对单个表的插入性能也有非常明显的提升，如果你的应用场景是日志型的，需要大批量的高并发入库，9.6就非常适合你。

LOCK改进，Partition the shared hash table freelist to reduce contention on multi-CPU-socket servers (Aleksander Alekseev)

本文将针对高并发的读，写，更新场景测试一下9.6和9.5的性能差异。

为了规避IO瓶颈的影响，体现9.6代码处理逻辑方面的改进，所有测试场景的数据均小于内存大小。

环境介绍

32核64HT, 512G, SSD, XFS。

全部在本地测试，避免网络的影响，但是本地测试有一个问题就是测试客户端也会占用一定的资源，特别是并发很高的时候，128个连接可能占用掉1/4的CPU资源。

如果网络允许，建议客户端使用另外的机器，比如我后来测试了客户端分离的情况，PG9.6 800个并发连接，tpc-b的查询依旧可以维持在110多万的TPS.

安装与配置

测试机器为同一主机。

1. OS配置

  1. # yum -y install coreutils glib2 lrzsz sysstat e4fsprogs xfsprogs ntp readline-devel zlib zlib-devel openssl openssl-devel pam-devel libxml2-devel libxslt-devel python-devel tcl-devel gcc make smartmontools flex bison perl perl-devel perl-ExtUtils* openldap openldap-devel
  3. # vi /etc/sysctl.conf
  4. # add by digoal.zhou
  5. fs.aio-max-nr = 1048576
  6. fs.file-max = 76724600
  7. kernel.core_pattern= /data01/corefiles/core_%e_%u_%t_%s.%p         
  8. # /data01/corefiles事先建好，权限777
  9. kernel.sem = 4096 2147483647 2147483646 512000    
  10. # 信号量, ipcs -l 或 -u 查看，每16个进程一组，每组信号量需要17个信号量。
  11. kernel.shmall = 107374182      
  12. # 所有共享内存段相加大小限制(建议内存的80%)
  13. kernel.shmmax = 274877906944   
  14. # 最大单个共享内存段大小(建议为内存一半), >9.2的版本已大幅降低共享内存的使用
  15. kernel.shmmni = 819200         
  16. # 一共能生成多少共享内存段，每个PG数据库集群至少2个共享内存段
  17. net.core.netdev_max_backlog = 10000
  18. net.core.rmem_default = 262144       
  19. # The default setting of the socket receive buffer in bytes.
  20. net.core.rmem_max = 4194304          
  21. # The maximum receive socket buffer size in bytes
  22. net.core.wmem_default = 262144       
  23. # The default setting (in bytes) of the socket send buffer.
  24. net.core.wmem_max = 4194304          
  25. # The maximum send socket buffer size in bytes.
  26. net.core.somaxconn = 4096
  27. net.ipv4.tcp_max_syn_backlog = 4096
  28. net.ipv4.tcp_keepalive_intvl = 20
  29. net.ipv4.tcp_keepalive_probes = 3
  30. net.ipv4.tcp_keepalive_time = 60
  31. net.ipv4.tcp_mem = 8388608 12582912 16777216
  32. net.ipv4.tcp_fin_timeout = 5
  33. net.ipv4.tcp_synack_retries = 2
  34. net.ipv4.tcp_syncookies = 1    
  35. # 开启SYN Cookies。当出现SYN等待队列溢出时，启用cookie来处理，可防范少量的SYN攻击
  36. net.ipv4.tcp_timestamps = 1    
  37. # 减少time_wait
  38. net.ipv4.tcp_tw_recycle = 0    
  39. # 如果=1则开启TCP连接中TIME-WAIT套接字的快速回收，但是NAT环境可能导致连接失败，建议服务端关闭它
  40. net.ipv4.tcp_tw_reuse = 1      
  41. # 开启重用。允许将TIME-WAIT套接字重新用于新的TCP连接
  42. net.ipv4.tcp_max_tw_buckets = 262144
  43. net.ipv4.tcp_rmem = 8192 87380 16777216
  44. net.ipv4.tcp_wmem = 8192 65536 16777216
  45. net.nf_conntrack_max = 1200000
  46. net.netfilter.nf_conntrack_max = 1200000
  47. vm.dirty_background_bytes = 409600000       
  48. #  系统脏页到达这个值，系统后台刷脏页调度进程 pdflush（或其他） 自动将(dirty_expire_centisecs/100）秒前的脏页刷到磁盘
  49. vm.dirty_expire_centisecs = 3000             
  50. #  比这个值老的脏页，将被刷到磁盘。3000表示30秒。
  51. vm.dirty_ratio = 95                          
  52. #  如果系统进程刷脏页太慢，使得系统脏页超过内存 95 % 时，则用户进程如果有写磁盘的操作（如fsync, fdatasync等调用），则需要主动把系统脏页刷出。
  53. #  有效防止用户进程刷脏页，在单机多实例，并且使用CGROUP限制单实例IOPS的情况下非常有效。  
  54. vm.dirty_writeback_centisecs = 100            
  55. #  pdflush（或其他）后台刷脏页进程的唤醒间隔， 100表示1秒。
  56. vm.extra_free_kbytes = 4096000
  57. vm.min_free_kbytes = 2097152
  58. vm.mmap_min_addr = 65536
  59. vm.overcommit_memory = 0     
  60. #  在分配内存时，允许少量over malloc, 如果设置为 1, 则认为总是有足够的内存，内存较少的测试环境可以使用 1 .  
  61. vm.overcommit_ratio = 90     
  62. #  当overcommit_memory = 2 时，用于参与计算允许指派的内存大小。
  63. vm.swappiness = 0            
  64. #  关闭交换分区
  65. vm.zone_reclaim_mode = 0     
  66. # 禁用 numa, 或者在vmlinux中禁止. 
  67. net.ipv4.ip_local_port_range = 40000 65535    
  68. # 本地自动分配的TCP, UDP端口号范围
  69. #  vm.nr_hugepages = 102352    
  70. #  建议shared buffer设置超过64GB时 使用大页，页大小 /proc/meminfo Hugepagesize
  72. # sysctl -p
  74. # vi /etc/security/limits.conf
  75. * soft    nofile  1024000
  76. * hard    nofile  1024000
  77. * soft    nproc   unlimited
  78. * hard    nproc   unlimited
  79. * soft    core    unlimited
  80. * hard    core    unlimited
  81. * soft    memlock unlimited
  82. * hard    memlock unlimited
  84. # rm -f /etc/security/limits.d/*

2. 数据库配置
安装

  1. $ wget https://ftp.postgresql.org/pub/source/v9.6.0/postgresql-9.6.0.tar.bz2
  2. $ wget https://ftp.postgresql.org/pub/source/v9.5.4/postgresql-9.5.4.tar.bz2
  4. $ tar -jxvf postgresql-9.5.4.tar.bz2
  5. $ tar -jxvf postgresql-9.6.0.tar.bz2
  7. $ cd ~/postgresql-9.6.0
  8. $ ./configure --prefix=/home/digoal/pgsql9.6.0
  9. $ make world -j 32
  10. $ make install-world -j 32
  12. $ cd ~/postgresql-9.5.4
  13. $ ./configure --prefix=/home/digoal/pgsql9.5
  14. $ make world -j 32
  15. $ make install-world -j 32
  17. $ vi ~/envpg96.sh
  18. export PS1="$USER@`/bin/hostname -s`-> "
  19. export PGPORT=5281
  20. export PGDATA=/data02/digoal/pg_root$PGPORT
  21. export LANG=en_US.utf8
  22. export PGHOME=/home/digoal/pgsql9.6.0
  23. export LD_LIBRARY_PATH=$PGHOME/lib:/lib64:/usr/lib64:/usr/local/lib64:/lib:/usr/lib:/usr/local/lib:$LD_LIBRARY_PATH
  24. export DATE=`date +"%Y%m%d%H%M"`
  25. export PATH=$PGHOME/bin:$PATH:.
  26. export MANPATH=$PGHOME/share/man:$MANPATH
  27. export PGHOST=$PGDATA
  28. export PGUSER=postgres
  29. export PGDATABASE=postgres
  30. alias rm='rm -i'
  31. alias ll='ls -lh'
  32. unalias vi
  34. $ vi ~/envpg95.sh
  35. export PS1="$USER@`/bin/hostname -s`-> "
  36. export PGPORT=5288
  37. export PGDATA=/data02/digoal/pg_root$PGPORT
  38. export LANG=en_US.utf8
  39. export PGHOME=/home/digoal/pgsql9.5
  40. export LD_LIBRARY_PATH=$PGHOME/lib:/lib64:/usr/lib64:/usr/local/lib64:/lib:/usr/lib:/usr/local/lib:$LD_LIBRARY_PATH
  41. export DATE=`date +"%Y%m%d%H%M"`
  42. export PATH=$PGHOME/bin:$PATH:.
  43. export MANPATH=$PGHOME/share/man:$MANPATH
  44. export PGHOST=$PGDATA
  45. export PGUSER=postgres
  46. export PGDATABASE=postgres
  47. alias rm='rm -i'
  48. alias ll='ls -lh'
  49. unalias vi
  51. $ df -h
  52. /dev/mapper/vgdata01-lv03
  53.                       4.0T  1.3T  2.8T  32% /u01
  54. /dev/mapper/vgdata01-lv04
  55.                       7.7T  899G  6.8T  12% /u02

初始化集群

  1. $ . ~/envpg96.sh
  2. $ initdb -D $PGDATA -E UTF8 --locale=C -U postgres -X /data01/digoal/pg_xlog$PGPORT
  4. $ . ~/envpg95.sh
  5. $ initdb -D $PGDATA -E UTF8 --locale=C -U postgres -X /data01/digoal/pg_xlog$PGPORT

配置数据库参数

  1. $ . ~/envpg96.sh
  2. $ cd $PGDATA
  5. $ vi postgresql.conf
  6. listen_addresses = '0.0.0.0'
  7. port = 5281
  8. max_connections = 800
  9. superuser_reserved_connections = 13
  10. unix_socket_directories = '.'
  11. unix_socket_permissions = 0700
  12. tcp_keepalives_idle = 60
  13. tcp_keepalives_interval = 10
  14. tcp_keepalives_count = 10
  15. shared_buffers = 128GB
  16. huge_pages = try
  17. maintenance_work_mem = 2GB
  18. dynamic_shared_memory_type = sysv
  19. vacuum_cost_delay = 0
  20. bgwriter_delay = 10ms
  21. bgwriter_lru_maxpages = 1000
  22. bgwriter_lru_multiplier = 10.0
  23. bgwriter_flush_after = 256
  24. max_worker_processes = 128
  25. max_parallel_workers_per_gather = 0
  26. old_snapshot_threshold = -1
  27. backend_flush_after = 0
  28. synchronous_commit = off
  29. full_page_writes = off
  30. wal_buffers = 1981MB
  31. wal_writer_delay = 10ms
  32. wal_writer_flush_after = 4MB
  33. checkpoint_timeout = 55min
  34. max_wal_size = 256GB
  35. checkpoint_flush_after = 1MB
  36. random_page_cost = 1.0
  37. effective_cache_size = 512GB
  38. constraint_exclusion = on  
  39. log_destination = 'csvlog'
  40. logging_collector = on
  41. log_checkpoints = on
  42. log_connections = on
  43. log_disconnections = on
  44. log_error_verbosity = verbose  
  45. log_timezone = 'PRC'
  46. autovacuum = on
  47. log_autovacuum_min_duration = 0
  48. autovacuum_max_workers = 8
  49. autovacuum_naptime = 10s
  50. autovacuum_vacuum_scale_factor = 0.02
  51. autovacuum_analyze_scale_factor = 0.01
  52. statement_timeout = 0
  53. lock_timeout = 0
  54. idle_in_transaction_session_timeout = 0
  55. gin_fuzzy_search_limit = 0
  56. gin_pending_list_limit = 4MB
  57. datestyle = 'iso, mdy'
  58. timezone = 'PRC'
  59. lc_messages = 'C'
  60. lc_monetary = 'C'
  61. lc_numeric = 'C'
  62. lc_time = 'C'
  63. default_text_search_config = 'pg_catalog.english'
  64. deadlock_timeout = 1s
  67. $ vi pg_hba.conf
  68. local   all             all                                     trust
  69. host    all             all             127.0.0.1/32            trust
  70. host    all             all             ::1/128                 trust
  71. host all all 0.0.0.0/0 md5
  74. $ . ~/envpg95.sh
  75. $ cd $PGDATA
  78. $ vi postgresql.conf
  79. listen_addresses = '0.0.0.0'
  80. port = 5288
  81. max_connections = 800
  82. superuser_reserved_connections = 13
  83. unix_socket_directories = '.'
  84. unix_socket_permissions = 0700
  85. tcp_keepalives_idle = 60
  86. tcp_keepalives_interval = 10
  87. tcp_keepalives_count = 10
  88. shared_buffers = 128GB
  89. huge_pages = try
  90. maintenance_work_mem = 2GB
  91. dynamic_shared_memory_type = posix
  92. vacuum_cost_delay = 0
  93. bgwriter_delay = 10ms
  94. bgwriter_lru_maxpages = 1000
  95. bgwriter_lru_multiplier = 10.0
  96. max_worker_processes = 128
  97. synchronous_commit = off
  98. full_page_writes = off
  99. wal_buffers = 1981MB
  100. wal_writer_delay = 10ms
  101. checkpoint_timeout = 55min
  102. max_wal_size = 256GB
  103. random_page_cost = 1.0
  104. effective_cache_size = 512GB
  105. constraint_exclusion = on  
  106. log_destination = 'csvlog'
  107. logging_collector = on
  108. log_checkpoints = on
  109. log_connections = on
  110. log_disconnections = on
  111. log_error_verbosity = verbose
  112. log_timezone = 'PRC'
  113. log_autovacuum_min_duration = 0
  114. autovacuum_max_workers = 8
  115. autovacuum_naptime = 10s
  116. autovacuum_vacuum_scale_factor = 0.02
  117. autovacuum_analyze_scale_factor = 0.01
  118. statement_timeout = 0
  119. lock_timeout = 0
  120. gin_fuzzy_search_limit = 0
  121. gin_pending_list_limit = 4MB
  122. datestyle = 'iso, mdy'
  123. timezone = 'PRC'
  124. lc_messages = 'C'
  125. lc_monetary = 'C'
  126. lc_numeric = 'C'
  127. lc_time = 'C'
  128. default_text_search_config = 'pg_catalog.english'
  129. deadlock_timeout = 1s
  132. $ vi pg_hba.conf
  133. local   all             all                                     trust
  134. host    all             all             127.0.0.1/32            trust
  135. host    all             all             ::1/128                 trust
  136. host all all 0.0.0.0/0 md5

启动数据库

  1. $ . ~/envpg96.sh
  2. $ pg_ctl start
  4. $ . ~/envpg95.sh
  5. $ pg_ctl start

测试时只启动一个数据库，防止干扰。

一、select based on PK only

环境准备

单表1亿数据量，基于PK的查询。

考察高并发下的代码优化能力。

SQL如下

  1. create table test(id int, info text, crt_time timestamp) with (autovacuum_freeze_max_age=1500000000, autovacuum_freeze_table_age=1400000000, autovacuum_multixact_freeze_max_age=1500000000, autovacuum_multixact_freeze_table_age=1400000000);
  2. insert into test select generate_series(1,100000000),md5(random()::text),clock_timestamp();
  3. set maintenance_work_mem='16GB';
  4. alter table test add constraint test_pkey primary key (id);
  5. vacuum analyze test;
  6. select * from test limit 10;
  7.  id |               info               |          crt_time          
  8. ----+----------------------------------+----------------------------
  9.   1 | 652802c64d630dfbde4770ed0d2a649c | 2016-10-02 15:38:12.866501
  10.   2 | c31d0e4ddd63618dbbb1c2a7932eae87 | 2016-10-02 15:38:12.866581
  11.   3 | f1689301bf26efd4050a88d50713ac66 | 2016-10-02 15:38:12.866586
  12.   4 | 155df78e2cd8f14291ddfd3f9179cde3 | 2016-10-02 15:38:12.866589
  13.   5 | 12aa2596dadb2af637bee07f05e78feb | 2016-10-02 15:38:12.866592
  14.   6 | 915f06af99501e629631b37f46f23816 | 2016-10-02 15:38:12.866595
  15.   7 | be79647d50351435b903c03a377e0ff5 | 2016-10-02 15:38:12.866597
  16.   8 | 676bedb18ffe2c7cc30a0d7ff081e7da | 2016-10-02 15:38:12.8666
  17.   9 | e7111e4c9f910ac00312f7a67ddbd162 | 2016-10-02 15:38:12.866602
  18.  10 | 22c6dd399e49663f3f14ce7634ff56d8 | 2016-10-02 15:38:12.866604
  19. (10 rows)

9.5

  1. $ vi test.sql
  2. \setrandom id 1 100000000
  3. select * from test where id=:id;
  5. $ vi bench.sh
  6. pgbench -M prepared -n -r -f ./test.sql -c 16 -j 16 -T 120
  7. pgbench -M prepared -n -r -f ./test.sql -c 32 -j 32 -T 120
  8. pgbench -M prepared -n -r -f ./test.sql -c 64 -j 64 -T 120
  9. pgbench -M prepared -n -r -f ./test.sql -c 72 -j 72 -T 120
  10. pgbench -M prepared -n -r -f ./test.sql -c 86 -j 86 -T 120
  11. pgbench -M prepared -n -r -f ./test.sql -c 96 -j 96 -T 120
  12. pgbench -M prepared -n -r -f ./test.sql -c 128 -j 128 -T 120
  13. pgbench -M prepared -n -r -f ./test.sql -c 192 -j 192 -T 120
  14. pgbench -M prepared -n -r -f ./test.sql -c 256 -j 256 -T 120
  16. $ . ./bench.sh

测试结果

并发数 , TPS

  1. 16 , 261687
  2. 32 , 514649
  3. 64 , 964129
  4. 72 , 946146
  5. 86 , 923699
  6. 96 , 931189
  7. 128 , 903589
  8. 192 , 891058
  9. 256 , 891150

9.6

  1. $ vi test.sql
  2. \set id random(1,100000000)
  3. select * from test where id=:id;
  5. $ vi bench.sh
  6. pgbench -M prepared -n -r -f ./test.sql -c 16 -j 16 -T 120
  7. pgbench -M prepared -n -r -f ./test.sql -c 32 -j 32 -T 120
  8. pgbench -M prepared -n -r -f ./test.sql -c 64 -j 64 -T 120
  9. pgbench -M prepared -n -r -f ./test.sql -c 72 -j 72 -T 120
  10. pgbench -M prepared -n -r -f ./test.sql -c 86 -j 86 -T 120
  11. pgbench -M prepared -n -r -f ./test.sql -c 96 -j 96 -T 120
  12. pgbench -M prepared -n -r -f ./test.sql -c 128 -j 128 -T 120
  13. pgbench -M prepared -n -r -f ./test.sql -c 192 -j 192 -T 120
  14. pgbench -M prepared -n -r -f ./test.sql -c 256 -j 256 -T 120
  16. $ . ./bench.sh

测试结果

并发数 , TPS

  1. 16 , 352524
  2. 32 , 611931
  3. 64 , 971911
  4. 72 , 994487
  5. 86 , 969640
  6. 96 , 970625
  7. 128 , 924109
  8. 192 , 893637
  9. 256 , 905555

对比

pic1

二、单表 update based on PK only

环境准备

单表1亿数据量，基于PK的更新。

考察高并发下的数据更新，autovacuum优化能力，XLOG优化能力。

SQL如下

  1. create table test(id int, info text, crt_time timestamp) with (autovacuum_freeze_max_age=1500000000, autovacuum_freeze_table_age=1400000000, autovacuum_multixact_freeze_max_age=1500000000, autovacuum_multixact_freeze_table_age=1400000000);;
  2. insert into test select generate_series(1,100000000),md5(random()::text),clock_timestamp();
  3. set maintenance_work_mem='16GB';
  4. alter table test add constraint test_pkey primary key (id);
  5. vacuum analyze test;
  6. select * from test limit 10;
  7.  id |               info               |          crt_time          
  8. ----+----------------------------------+----------------------------
  9.   1 | 652802c64d630dfbde4770ed0d2a649c | 2016-10-02 15:38:12.866501
  10.   2 | c31d0e4ddd63618dbbb1c2a7932eae87 | 2016-10-02 15:38:12.866581
  11.   3 | f1689301bf26efd4050a88d50713ac66 | 2016-10-02 15:38:12.866586
  12.   4 | 155df78e2cd8f14291ddfd3f9179cde3 | 2016-10-02 15:38:12.866589
  13.   5 | 12aa2596dadb2af637bee07f05e78feb | 2016-10-02 15:38:12.866592
  14.   6 | 915f06af99501e629631b37f46f23816 | 2016-10-02 15:38:12.866595
  15.   7 | be79647d50351435b903c03a377e0ff5 | 2016-10-02 15:38:12.866597
  16.   8 | 676bedb18ffe2c7cc30a0d7ff081e7da | 2016-10-02 15:38:12.8666
  17.   9 | e7111e4c9f910ac00312f7a67ddbd162 | 2016-10-02 15:38:12.866602
  18.  10 | 22c6dd399e49663f3f14ce7634ff56d8 | 2016-10-02 15:38:12.866604
  19. (10 rows)

9.5

  1. $ vi test.sql
  2. \setrandom id 1 100000000
  3. update test set crt_time=now() where id=:id;
  5. $ vi bench.sh
  6. pgbench -M prepared -n -r -f ./test.sql -c 16 -j 16 -T 120
  7. pgbench -M prepared -n -r -f ./test.sql -c 32 -j 32 -T 120
  8. pgbench -M prepared -n -r -f ./test.sql -c 64 -j 64 -T 120
  9. pgbench -M prepared -n -r -f ./test.sql -c 72 -j 72 -T 120
  10. pgbench -M prepared -n -r -f ./test.sql -c 86 -j 86 -T 120
  11. pgbench -M prepared -n -r -f ./test.sql -c 96 -j 96 -T 120
  12. pgbench -M prepared -n -r -f ./test.sql -c 128 -j 128 -T 120
  13. pgbench -M prepared -n -r -f ./test.sql -c 192 -j 192 -T 120
  14. pgbench -M prepared -n -r -f ./test.sql -c 256 -j 256 -T 120
  16. $ . ./bench.sh

并发数 , TPS

  1. 16 , 160502
  2. 32 , 202785
  3. 64 , 146669
  4. 72 , 136701
  5. 86 , 124060
  6. 96 , 116345
  7. 128 , 100642
  8. 192 , 76714
  9. 256 , 57945

9.6

  1. $ vi test.sql
  2. \set id random(1,100000000)
  3. update test set crt_time=now() where id=:id;
  5. $ vi bench.sh
  6. pgbench -M prepared -n -r -f ./test.sql -c 16 -j 16 -T 120
  7. pgbench -M prepared -n -r -f ./test.sql -c 32 -j 32 -T 120
  8. pgbench -M prepared -n -r -f ./test.sql -c 64 -j 64 -T 120
  9. pgbench -M prepared -n -r -f ./test.sql -c 72 -j 72 -T 120
  10. pgbench -M prepared -n -r -f ./test.sql -c 86 -j 86 -T 120
  11. pgbench -M prepared -n -r -f ./test.sql -c 96 -j 96 -T 120
  12. pgbench -M prepared -n -r -f ./test.sql -c 128 -j 128 -T 120
  13. pgbench -M prepared -n -r -f ./test.sql -c 192 -j 192 -T 120
  14. pgbench -M prepared -n -r -f ./test.sql -c 256 -j 256 -T 120
  16. $ . ./bench.sh

并发数 , TPS

  1. 16 , 216928
  2. 32 , 289555
  3. 64 , 249844
  4. 72 , 233400
  5. 86 , 214760
  6. 96 , 203196
  7. 128 , 178891
  8. 192 , 152073
  9. 256 , 129707

对比

pic2

三、单表 autocommit 单条 insert only

环境准备

一张空表，22个字段，每行约201字节，包含两个索引。

采用autocommit的模式，每个连接每个事务插入一条记录。

考察高并发下的数据插入，数据块扩展能力，XLOG优化能力。

SQL如下

  1. create table test(id serial8, c1 int8 default 0, c2 int8 default 0, c3 int8 default 0, c4 int8 default 0, c5 int8 default 0, c6 int8 default 0, c7 int8 default 0, c8 int8 default 0, c9 int8 default 0, c10 int8 default 0, c11 int8 default 0, c12 int8 default 0, c13 int8 default 0, c14 int8 default 0, c15 int8 default 0, c16 int8 default 0, c17 int8 default 0, c18 int8 default 0, c19 int8 default 0, c20 int8 default 0, crt_time timestamptz) with (autovacuum_enabled=off, autovacuum_freeze_max_age=1500000000, autovacuum_freeze_table_age=1400000000, autovacuum_multixact_freeze_max_age=1500000000, autovacuum_multixact_freeze_table_age=1400000000); 
  2. alter sequence test_id_seq cache 100000;
  3. create index idx_test_1 on test using brin(id);
  4. create index idx_test_2 on test using brin(crt_time);

测试脚本如下

  1. $ vi test.sql
  2. insert into test(crt_time) values(now());
  4. $ vi bench.sh
  5. pgbench -M prepared -n -r -f ./test.sql -c 16 -j 16 -T 120
  6. pgbench -M prepared -n -r -f ./test.sql -c 32 -j 32 -T 120
  7. pgbench -M prepared -n -r -f ./test.sql -c 64 -j 64 -T 120
  8. pgbench -M prepared -n -r -f ./test.sql -c 72 -j 72 -T 120
  9. pgbench -M prepared -n -r -f ./test.sql -c 86 -j 86 -T 120
  10. pgbench -M prepared -n -r -f ./test.sql -c 96 -j 96 -T 120
  11. pgbench -M prepared -n -r -f ./test.sql -c 128 -j 128 -T 120
  12. pgbench -M prepared -n -r -f ./test.sql -c 192 -j 192 -T 120
  13. pgbench -M prepared -n -r -f ./test.sql -c 256 -j 256 -T 120
  15. $ . ./bench.sh

9.5

并发数 , TPS

  1. 16 , 234043
  2. 32 , 263893
  3. 64 , 208993
  4. 72 , 199966
  5. 86 , 188826
  6. 96 , 182672
  7. 128 , 164270
  8. 192 , 130384
  9. 256 , 104563

9.6

并发数 , TPS

  1. 16 , 268877
  2. 32 , 313320
  3. 64 , 324775
  4. 72 , 318060
  5. 86 , 307001
  6. 96 , 296028
  7. 128 , 256317
  8. 192 , 202902
  9. 256 , 154469

对比

pic3

四、单表 autocommit 批量 insert only

环境准备

批量插入，考察的同样是高并发处理单表时XLOG的优化能力，数据文件的扩展优化能力。

测试脚本如下

一次插入400条记录。

  1. $ vi test.sql
  2. 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()),(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()),(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());
  5. $ . ./bench.sh

9.5

并发数 , TPS

  1. 16 , 2875
  2. 32 , 2752
  3. 64 , 2534
  4. 72 , 2473
  5. 86 , 2424
  6. 96 , 2372
  7. 128 , 2362
  8. 192 , 2283
  9. 256 , 2140

9.6

并发数 , TPS

  1. 16 , 3450
  2. 32 , 3363
  3. 64 , 2905
  4. 72 , 2792
  5. 86 , 3155
  6. 96 , 3320
  7. 128 , 2992
  8. 192 , 3152
  9. 256 , 3070

对比

pic4

五、多表 autocommit 单条 insert only

环境准备

每个连接对应一张空表，22个字段，每行约201字节，包含两个索引。

采用autocommit的模式，每个连接每个事务插入一条记录。

考察高并发下的数据插入，XLOG优化能力。

与单表不同，因为没有单表的文件扩展并发要求，所以不考察数据块扩展能力。

SQL如下

  1. create table test(id serial8, c1 int8 default 0, c2 int8 default 0, c3 int8 default 0, c4 int8 default 0, c5 int8 default 0, c6 int8 default 0, c7 int8 default 0, c8 int8 default 0, c9 int8 default 0, c10 int8 default 0, c11 int8 default 0, c12 int8 default 0, c13 int8 default 0, c14 int8 default 0, c15 int8 default 0, c16 int8 default 0, c17 int8 default 0, c18 int8 default 0, c19 int8 default 0, c20 int8 default 0, crt_time timestamptz) with (autovacuum_enabled=off, autovacuum_freeze_max_age=1500000000, autovacuum_freeze_table_age=1400000000, autovacuum_multixact_freeze_max_age=1500000000, autovacuum_multixact_freeze_table_age=1400000000);  
  2. alter sequence test_id_seq cache 100000;
  3. create index idx_test_1 on test using brin(id);
  4. create index idx_test_2 on test using brin(crt_time);

批量创建测试表，测试脚本

  1. for ((i=1;i<=256;i++)); do psql -c "create table test$i(like test including all) with (autovacuum_enabled=off, autovacuum_freeze_max_age=1500000000, autovacuum_freeze_table_age=1400000000, autovacuum_multixact_freeze_max_age=1500000000, autovacuum_multixact_freeze_table_age=1400000000)"; done
  3. for ((i=1;i<=256;i++)); do echo "insert into test$i(crt_time) values(now());" > ~/test$i.sql; done

测试脚本如下

  1. $ vi bench.sh
  2. for ((i=1;i<=16;i++)); do psql -c "truncate test$i"; done
  3. psql -c "checkpoint;"
  4. for ((i=1;i<=16;i++)); do pgbench -M prepared -n -r -f ./test$i.sql -c 1 -j 1 -T 120 >/tmp/test_16_$i.log & done
  5. sleep 130
  6. for ((i=1;i<=32;i++)); do psql -c "truncate test$i"; done
  7. psql -c "checkpoint;"
  8. for ((i=1;i<=32;i++)); do pgbench -M prepared -n -r -f ./test$i.sql -c 1 -j 1 -T 120 >/tmp/test_32_$i.log & done
  9. sleep 130
  10. for ((i=1;i<=64;i++)); do psql -c "truncate test$i"; done
  11. psql -c "checkpoint;"
  12. for ((i=1;i<=64;i++)); do pgbench -M prepared -n -r -f ./test$i.sql -c 1 -j 1 -T 120 >/tmp/test_64_$i.log & done
  13. sleep 130
  14. for ((i=1;i<=72;i++)); do psql -c "truncate test$i"; done
  15. psql -c "checkpoint;"
  16. for ((i=1;i<=72;i++)); do pgbench -M prepared -n -r -f ./test$i.sql -c 1 -j 1 -T 120 >/tmp/test_72_$i.log & done
  17. sleep 130
  18. for ((i=1;i<=86;i++)); do psql -c "truncate test$i"; done
  19. psql -c "checkpoint;"
  20. for ((i=1;i<=86;i++)); do pgbench -M prepared -n -r -f ./test$i.sql -c 1 -j 1 -T 120 >/tmp/test_86_$i.log & done
  21. sleep 130
  22. for ((i=1;i<=96;i++)); do psql -c "truncate test$i"; done
  23. psql -c "checkpoint;"
  24. for ((i=1;i<=96;i++)); do pgbench -M prepared -n -r -f ./test$i.sql -c 1 -j 1 -T 120 >/tmp/test_96_$i.log & done
  25. sleep 130
  26. for ((i=1;i<=128;i++)); do psql -c "truncate test$i"; done
  27. psql -c "checkpoint;"
  28. for ((i=1;i<=128;i++)); do pgbench -M prepared -n -r -f ./test$i.sql -c 1 -j 1 -T 120 >/tmp/test_128_$i.log & done
  29. sleep 130
  30. for ((i=1;i<=192;i++)); do psql -c "truncate test$i"; done
  31. psql -c "checkpoint;"
  32. for ((i=1;i<=192;i++)); do pgbench -M prepared -n -r -f ./test$i.sql -c 1 -j 1 -T 120 >/tmp/test_192_$i.log & done
  33. sleep 130
  34. for ((i=1;i<=256;i++)); do psql -c "truncate test$i"; done
  35. psql -c "checkpoint;"
  36. for ((i=1;i<=256;i++)); do pgbench -M prepared -n -r -f ./test$i.sql -c 1 -j 1 -T 120 >/tmp/test_256_$i.log & done
  37. sleep 130
  40. $ . ./bench.sh

统计

  1. $ vi res.sh
  3. x=0; for ((i=1;i<=16;i++)); do y=`cat /tmp/test_16_$i.log | grep "excluding connections establishing" | awk '{print $3}' | awk -F "." '{print $1}'`; x=$(($x+$y)); done; echo "16 , $x"
  4. x=0; for ((i=1;i<=32;i++)); do y=`cat /tmp/test_32_$i.log | grep "excluding connections establishing" | awk '{print $3}' | awk -F "." '{print $1}'`; x=$(($x+$y)); done; echo "32 , $x"
  5. x=0; for ((i=1;i<=64;i++)); do y=`cat /tmp/test_64_$i.log | grep "excluding connections establishing" | awk '{print $3}' | awk -F "." '{print $1}'`; x=$(($x+$y)); done; echo "64 , $x"
  6. x=0; for ((i=1;i<=72;i++)); do y=`cat /tmp/test_72_$i.log | grep "excluding connections establishing" | awk '{print $3}' | awk -F "." '{print $1}'`; x=$(($x+$y)); done; echo "72 , $x"
  7. x=0; for ((i=1;i<=86;i++)); do y=`cat /tmp/test_86_$i.log | grep "excluding connections establishing" | awk '{print $3}' | awk -F "." '{print $1}'`; x=$(($x+$y)); done; echo "86 , $x"
  8. x=0; for ((i=1;i<=96;i++)); do y=`cat /tmp/test_96_$i.log | grep "excluding connections establishing" | awk '{print $3}' | awk -F "." '{print $1}'`; x=$(($x+$y)); done; echo "96 , $x"
  9. x=0; for ((i=1;i<=128;i++)); do y=`cat /tmp/test_128_$i.log | grep "excluding connections establishing" | awk '{print $3}' | awk -F "." '{print $1}'`; x=$(($x+$y)); done; echo "128 , $x"
  10. x=0; for ((i=1;i<=192;i++)); do y=`cat /tmp/test_192_$i.log | grep "excluding connections establishing" | awk '{print $3}' | awk -F "." '{print $1}'`; x=$(($x+$y)); done; echo "192 , $x"
  11. x=0; for ((i=1;i<=256;i++)); do y=`cat /tmp/test_256_$i.log | grep "excluding connections establishing" | awk '{print $3}' | awk -F "." '{print $1}'`; x=$(($x+$y)); done; echo "256 , $x"
  14. $ . ./res.sh

9.5

并发数 , TPS

  1. 16 , 225198
  2. 32 , 280587
  3. 64 , 222368
  4. 72 , 213024
  5. 86 , 199209
  6. 96 , 190801
  7. 128 , 167913
  8. 192 , 131405
  9. 256 , 102913

9.6

并发数 , TPS

  1. 16 , 288706
  2. 32 , 351340
  3. 64 , 382612
  4. 72 , 377392
  5. 86 , 362909
  6. 96 , 334932
  7. 128 , 279157
  8. 192 , 200568
  9. 256 , 152104

对比

pic5

六、多表 autocommit 批量 insert only

环境准备

批量插入，考察的同样是高并发处理单表时XLOG的优化能力，数据文件的扩展优化能力。

测试脚本如下

一次插入400条记录。

  1. for ((i=1;i<=256;i++)); do echo "insert into test$i(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()),(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()),(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());" > ~/test$i.sql; done
  4. $ . ./bench.sh

统计

  1. $ . ./res.sh

9.5

并发数 , TPS

  1. 16 , 5693
  2. 32 , 5767
  3. 64 , 5297
  4. 72 , 4073
  5. 86 , 5374
  6. 96 , 4978
  7. 128 , 5438
  8. 192 , 5247
  9. 256 , 5376

9.6

并发数 , TPS

  1. 16 , 6007
  2. 32 , 6120
  3. 64 , 5289
  4. 72 , 5501
  5. 86 , 5503
  6. 96 , 5605
  7. 128 , 5537
  8. 192 , 5467
  9. 256 , 5376

对比

pic6

小结

PostgreSQL 9.6的锁控制能力又有比较大的进步，在WAL的高并发管理，获取快照，扩展数据文件等方面都有较大改进，相比9.5在scale-up的扩展能力上又上了一个新的台阶，在高并发的读，插入，更新场景，都有非常明显的性能提升。

结合9.6的多核并行计算，可以适合高并发的TP场景，又能在业务低谷时充分发挥硬件能力，处理AP的报表和分析需求，完成业务对TP+AP的混合需求。

对于3,4,5,6的测试CASE，由于是批量入库，可以关闭测试表的autovacuum，达到更好的性能。

现在的CPU一直在往多核的方向发展，32核已经是非常普遍的配置，多的甚至可以达到上千核。

使用PostgreSQL可以更好的发挥硬件的性能，虽然PostgreSQL已经在内核层面支持sharding了，但是使用单机能解决的场景，不推荐使用sharding。

目前sharding对应用开发的限制还比较多，比如大多数sharding技术需要解决几个痛点：

分布式事务的控制，跨库JOIN，全局一致性，全局约束，数据倾斜，扩容，备份，容灾，迁移，确保全局一致性的高可用技术。等等一系列需要考虑的问题。

原文：http://mysql.taobao.org/monthly/2016/11/06/