承接上一篇XFS性能优化。
本文是对比XFS优化前后的PostgreSQL TPC-B性能。
配置PostgreSQL
数据库参数
listen_addresses='0.0.0.0'
port=1921
max_connections=1000
unix_socket_directories='.'
shared_buffers=32GB
maintenance_work_mem=512MB
autovacuum_work_mem=512MB
dynamic_shared_memory_type=posix
bgwriter_delay=10ms
wal_level=hot_standby
synchronous_commit=off
full_page_writes=off
wal_buffers=16MB
wal_writer_delay=10ms
max_wal_size=32GB
max_wal_senders=10
max_replication_slots=10
hot_standby=on
wal_receiver_status_interval=1s
hot_standby_feedback=on
random_page_cost=1.0
effective_cache_size=256GB
log_destination='csvlog'
logging_collector=on
log_checkpoints=on
log_connections=on
log_disconnections=on
log_error_verbosity=verbose
log_timezone='PRC'
datestyle='iso,
timezone='PRC'
lc_messages='C'
lc_monetary='C'
lc_numeric='C'
lc_time='C'
default_text_search_config='pg_catalog.english'
$PGDATA和pg_xlog都放在XFS文件系统中。
生成测试数据
生成5亿测试数据
pgbench -i -s 5000
初始化性能对比
XFS未优化时的统计信息如下
pgbench初始化测试数据时的磁盘iostat
IO落在一块盘上
平均IO响应0.36毫秒，平均IO等待0.87毫秒
avg-cpu: %user %nice %system %iowait %steal %idle
3.60 0.00 1.35 0.19 0.00 94.87
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
sda 0.00 1.00 0.00 4.00 0.00 40.00 10.00 0.00 0.00 0.00 0.00
sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfa 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfc 0.00 0.00 0.00 250.00 0.00 206752.00 827.01 0.22 0.86 0.36 9.10
dm-0 0.00 0.00 0.00 250.00 0.00 206752.00 827.01 0.22 0.87 0.36 9.10
vacuum 时的iostat
IO落在一块盘上
平均IO响应0.28毫秒，平均IO等待26.56毫秒
avg-cpu: %user %nice %system %iowait %steal %idle
1.88 0.00 4.77 0.00 0.00 93.35
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
sda 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfa 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfb 0.00 0.00 0.00 3211.00 0.00 3280888.00 1021.77 86.20 26.53 0.28 90.20
dfc 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dm-0 0.00 0.00 0.00 3210.00 0.00 3279864.00 1021.76 86.21 26.56 0.28 90.20
XFS未优化后的统计信息如下
pgbench初始化测试数据时的磁盘iostat
IO分摊到多快盘上
平均IO响应0.00毫秒，平均IO等待0.01毫秒
avg-cpu: %user %nice %system %iowait %steal %idle
3.57 0.00 2.50 0.03 0.00 93.90
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
sda 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfa 0.00 0.00 0.00 9262.00 0.00 148096.00 15.99 0.11 0.01 0.01 8.60
dfb 0.00 0.00 0.00 9243.00 0.00 147888.00 16.00 0.11 0.01 0.01 8.80
dfc 0.00 0.00 0.00 9256.00 0.00 148096.00 16.00 0.17 0.02 0.01 9.00
dm-0 0.00 0.00 0.00 27761.00 0.00 444080.00 16.00 0.42 0.01 0.00 9.30
vacuum 时的iostat
IO分摊到多快盘上
平均IO响应0.01毫秒，平均IO等待0.02毫秒
avg-cpu: %user %nice %system %iowait %steal %idle
2.09 0.00 13.57 0.10 0.00 84.25
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
sda 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfa 0.00 0.00 0.00 52311.00 0.00 836976.00 16.00 0.70 0.01 0.01 63.50
dfb 0.00 0.00 0.00 52313.00 0.00 837008.00 16.00 0.74 0.01 0.01 66.90
dfc 0.00 0.00 0.00 52312.00 0.00 836984.00 16.00 1.11 0.02 0.02 94.30
dm-0 0.00 0.00 0.00 156936.00 0.00 2510952.00 16.00 2.90 0.02 0.01 97.40
压测性能对比
压测命令
nohup pgbench -M prepared -n -r -P 1 -c 96 -j 96 -T 3600 >./bench.log 2>&1 &
XFS优化前性能指标
XFS优化前的IOSTAT
未发生检查点时
IO落到一个块设备
平均IO响应0.04毫秒，平均IO等待0.1毫秒
avg-cpu: %user %nice %system %iowait %steal %idle
73.16 0.00 22.17 0.06 0.00 4.61
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
sda 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfa 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfb 0.00 0.00 0.00 922.00 0.00 12456.00 13.51 0.00 0.00 0.00 0.30
dfc 0.00 0.00 0.00 3955.00 0.00 131776.00 33.32 0.44 0.11 0.05 18.20
dm-0 0.00 0.00 0.00 4877.00 0.00 144232.00 29.57 0.47 0.10 0.04 21.00
发生检查点时
IO落到一个块设备
平均IO响应0.03毫秒，平均IO等待0.6毫秒
avg-cpu: %user %nice %system %iowait %steal %idle
72.67 0.00 25.48 0.00 0.00 1.85
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
sda 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfa 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfb 0.00 0.00 0.00 32539.00 0.00 3184272.00 97.86 20.65 0.63 0.03 100.20
dfc 0.00 0.00 0.00 2566.00 0.00 84464.00 32.92 0.35 0.14 0.09 22.90
dm-0 0.00 0.00 0.00 35104.00 0.00 3268672.00 93.11 21.06 0.60 0.03 100.10
tps
transaction type: TPC-B (sort of)
scaling factor: 5000
query mode: prepared
number of clients: 96
number of threads: 96
duration: 3600 s
number of transactions actually processed: 34796111
latency average: 9.929 ms
latency stddev: 14.829 ms
tps = 9664.902474 (including connections establishing)
tps = 9665.028796 (excluding connections establishing)
statement latencies in milliseconds:
0.006275 \set nbranches 1 * :scale
0.001541 \set ntellers 10 * :scale
0.001213 \set naccounts 100000 * :scale
0.002230 \setrandom aid 1 :naccounts
0.001708 \setrandom bid 1 :nbranches
0.001497 \setrandom tid 1 :ntellers
0.001460 \setrandom delta -5000 5000
1.284175 BEGIN;
1.139272 UPDATE pgbench_accounts SET abalance = abalance + :delta WHERE aid = :aid;
1.012388 SELECT abalance FROM pgbench_accounts WHERE aid = :aid;
1.682350 UPDATE pgbench_tellers SET tbalance = tbalance + :delta WHERE tid = :tid;
2.705569 UPDATE pgbench_branches SET bbalance = bbalance + :delta WHERE bid = :bid;
1.058834 INSERT INTO pgbench_history (tid, bid, aid, delta, mtime) VALUES (:tid, :bid, :aid, :delta, CURRENT_TIMESTAMP);
1.015118 END;
XFS优化后性能指标
未发生检查点时
IO分摊到多快盘上
平均IO响应0.02毫秒，平均IO等待0.02毫秒
avg-cpu: %user %nice %system %iowait %steal %idle
65.34 0.00 26.94 0.00 0.00 7.72
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
sda 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfa 0.00 0.00 0.00 1428.00 0.00 22832.00 15.99 0.02 0.02 0.02 2.40
dfb 0.00 0.00 0.00 1417.00 0.00 22672.00 16.00 0.03 0.02 0.02 2.50
dfc 0.00 0.00 0.00 1417.00 0.00 22672.00 16.00 0.04 0.02 0.02 3.50
dm-0 0.00 0.00 0.00 4262.00 0.00 68176.00 16.00 0.10 0.02 0.02 9.50
检查点发生时
IO分摊到多快盘上
平均IO响应0.01毫秒，平均IO等待0.02毫秒
avg-cpu: %user %nice %system %iowait %steal %idle
58.82 0.00 35.25 0.00 0.00 5.93
Device: rrqm/s wrqm/s r/s w/s rsec/s wsec/s avgrq-sz avgqu-sz await svctm %util
sda 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dfa 0.00 0.00 0.00 52709.00 0.00 843272.00 16.00 1.12 0.02 0.01 72.30
dfb 0.00 0.00 0.00 52632.00 0.00 842048.00 16.00 0.95 0.02 0.01 74.10
dfc 0.00 0.00 0.00 52805.00 0.00 844832.00 16.00 1.32 0.02 0.02 89.90
dm-0 0.00 0.00 0.00 158143.00 0.00 2530120.00 16.00 3.72 0.02 0.01 99.50
tps
transaction type: TPC-B (sort of)
scaling factor: 5000
query mode: prepared
number of clients: 96
number of threads: 96
duration: 3600 s
number of transactions actually processed: 144901132
latency average: 2.381 ms
latency stddev: 1.571 ms
tps = 40248.632845 (including connections establishing)
tps = 40249.214064 (excluding connections establishing)
statement latencies in milliseconds:
0.006825 \set nbranches 1 * :scale
0.002457 \set ntellers 10 * :scale
0.001856 \set naccounts 100000 * :scale
0.003474 \setrandom aid 1 :naccounts
0.002448 \setrandom bid 1 :nbranches
0.002864 \setrandom tid 1 :ntellers
0.003144 \setrandom delta -5000 5000
0.220207 BEGIN;
0.372330 UPDATE pgbench_accounts SET abalance = abalance + :delta WHERE aid = :aid;
0.247961 SELECT abalance FROM pgbench_accounts WHERE aid = :aid;
0.301182 UPDATE pgbench_tellers SET tbalance = tbalance + :delta WHERE tid = :tid;
0.305694 UPDATE pgbench_branches SET bbalance = bbalance + :delta WHERE bid = :bid;
0.247623 INSERT INTO pgbench_history (tid, bid, aid, delta, mtime) VALUES (:tid, :bid, :aid, :delta, CURRENT_TIMESTAMP);
0.645872 END;
tps曲线
小结
经过XFS优化，单次IO响应延迟始终保持在0.01毫秒左右，性能表现平稳。
使用了9000个allocation group, fsync时，并行度更高，所以对TPS影响更低了。
