Proxying MySQL, benchmarking on production hardware
Fri, Feb 26, 2021When connecting to a remote MySQL server there can be a pretty big performance impact, just due the fact we are using TCP to transfer our data. Where connecting over a socket is usually blazingly fast, introducing TCP to do that might give us a performance penalty on our application. We have seen some information on local tests, but how do these behave in the real world. So we will need to test how these sysbench tests will behave on actual production hardware. How much performance drop we see when we actually connect remotely to MySQL and is the difference between the proxies still pretty big.
Goal
We want to find out if we can move MySQL to separate machines. If we can, that will make the overall maintenance of machines in our cluster easier since we can make those more single purpose.
What will we test
For MySQL we want to test both 5.7 and 8.0, we can’t keep running MySQL 5.x forever :).
And as proxies we will use the same ones we used in the local tests. We will use the configurations as they can be found in the playground, only with different ip adresses and the user / password combinations will be different, but these don’t impose a relevant change to the test.
The test itself will be run with sysbench.
#!/usr/bin/env bash
sysbench \
/usr/share/sysbench/oltp_read_only.lua \
--threads="2" \
--tables=10 \
--table-size=1000000 \
--report-interval=5 \
--rand-type=pareto \
--forced-shutdown=1 \
--time=300 \
--events=0 \
--point-selects=25 \
--range_size=5 \
--skip_trx=on \
--percentile=95 \
--mysql-host=mysql.to.test \
--mysql-port=3306 \
--mysql-user=bench \
--mysql-password=our-fancy-bench-password \
--mysql-db=bench \
--mysql-storage-engine=INNODB \
run
We will run with only 2 threads so we always compare the same information.
In our results we will be most interested in the number of transactions / queries per second and what the 95% latency is during that time.
MySQL 5.7 vs 8.0 on local machine
As a baseline we already want to know how both these versions compare when we
run sysbench on the local machine over TCP.
Results for MySQL 5.7
SQL statistics:
queries performed:
read: 9553412
write: 0
other: 0
total: 9553412
transactions: 329428 (1098.07 per sec.)
queries: 9553412 (31844.15 per sec.)
ignored errors: 0 (0.00 per sec.)
reconnects: 0 (0.00 per sec.)
General statistics:
total time: 300.0039s
total number of events: 329428
Latency (ms):
min: 1.18
avg: 1.82
max: 17.65
95th percentile: 2.03
sum: 599617.66
Threads fairness:
events (avg/stddev): 164714.0000/9.00
execution time (avg/stddev): 299.8088/0.00
Results for MySQL 8.0
SQL statistics:
queries performed:
read: 8120406
write: 0
other: 0
total: 8120406
transactions: 280014 (933.37 per sec.)
queries: 8120406 (27067.71 per sec.)
ignored errors: 0 (0.00 per sec.)
reconnects: 0 (0.00 per sec.)
General statistics:
total time: 300.0021s
total number of events: 280014
Latency (ms):
min: 1.35
avg: 2.14
max: 18.71
95th percentile: 2.61
sum: 599620.24
Threads fairness:
events (avg/stddev): 140007.0000/26.00
execution time (avg/stddev): 299.8101/0.00
Summary:
| type | transactions / sec | queries / sec | 95% latency (ms) |
|---|---|---|---|
| MySQL 5.7 | 1098.07 | 31844.15 | 2.03 |
| MySQL 8.0 | 933.37 | 27067.71 | 2.61 |
So we see already that MySQL 8.0 only gives us 85% of the troughput with the same configuration. Also the percieved latency increases when we start using MySQL 8.0.
MySQL 5.7 vs 8.0 from remote with and without proxies
Let us see how the different proxies / remote connections behave.
Results direct connection to MySQL 5.7
SQL statistics:
queries performed:
read: 3452044
write: 0
other: 0
total: 3452044
transactions: 119036 (396.77 per sec.)
queries: 3452044 (11506.45 per sec.)
ignored errors: 0 (0.00 per sec.)
reconnects: 0 (0.00 per sec.)
General statistics:
total time: 300.0082s
total number of events: 119036
Latency (ms):
min: 3.33
avg: 5.04
max: 235.47
95th percentile: 6.21
sum: 599827.35
Threads fairness:
events (avg/stddev): 59518.0000/164.00
execution time (avg/stddev): 299.9137/0.00
Results direct connection to MySQL 8.0
SQL statistics:
queries performed:
read: 3113846
write: 0
other: 0
total: 3113846
transactions: 107374 (357.90 per sec.)
queries: 3113846 (10379.20 per sec.)
ignored errors: 0 (0.00 per sec.)
reconnects: 0 (0.00 per sec.)
General statistics:
total time: 300.0068s
total number of events: 107374
Latency (ms):
min: 3.53
avg: 5.59
max: 235.52
95th percentile: 6.79
sum: 599819.55
Threads fairness:
events (avg/stddev): 53687.0000/58.00
execution time (avg/stddev): 299.9098/0.00
Results Haproxy connection to MySQL 5.7
SQL statistics:
queries performed:
read: 2818568
write: 0
other: 0
total: 2818568
transactions: 97192 (323.97 per sec.)
queries: 2818568 (9395.09 per sec.)
ignored errors: 0 (0.00 per sec.)
reconnects: 0 (0.00 per sec.)
General statistics:
total time: 300.0031s
total number of events: 97192
Latency (ms):
min: 4.46
avg: 6.17
max: 226.03
95th percentile: 7.43
sum: 599824.61
Threads fairness:
events (avg/stddev): 48596.0000/90.00
execution time (avg/stddev): 299.9123/0.00
Results Haproxy connection to MySQL 8.0
SQL statistics:
queries performed:
read: 2792990
write: 0
other: 0
total: 2792990
transactions: 96310 (321.03 per sec.)
queries: 2792990 (9309.79 per sec.)
ignored errors: 0 (0.00 per sec.)
reconnects: 0 (0.00 per sec.)
General statistics:
total time: 300.0044s
total number of events: 96310
Latency (ms):
min: 4.16
avg: 6.23
max: 219.98
95th percentile: 7.43
sum: 599839.42
Threads fairness:
events (avg/stddev): 48155.0000/79.00
execution time (avg/stddev): 299.9197/0.00
Results Traefik connection to MySQL 5.7
SQL statistics:
queries performed:
read: 2846756
write: 0
other: 0
total: 2846756
transactions: 98164 (327.21 per sec.)
queries: 2846756 (9488.97 per sec.)
ignored errors: 0 (0.00 per sec.)
reconnects: 0 (0.00 per sec.)
General statistics:
total time: 300.0054s
total number of events: 98164
Latency (ms):
min: 4.45
avg: 6.11
max: 299.99
95th percentile: 7.43
sum: 599828.18
Threads fairness:
events (avg/stddev): 49082.0000/178.00
execution time (avg/stddev): 299.9141/0.00
Results Traefik connection to MySQL 8.0
SQL statistics:
queries performed:
read: 2613973
write: 0
other: 0
total: 2613973
transactions: 90137 (300.45 per sec.)
queries: 2613973 (8712.97 per sec.)
ignored errors: 0 (0.00 per sec.)
reconnects: 0 (0.00 per sec.)
General statistics:
total time: 300.0079s
total number of events: 90137
Latency (ms):
min: 4.45
avg: 6.66
max: 201.54
95th percentile: 7.84
sum: 599871.02
Threads fairness:
events (avg/stddev): 45068.5000/5.50
execution time (avg/stddev): 299.9355/0.00
Results ProxySQL connection to MySQL 5.7
SQL statistics:
queries performed:
read: 2004335
write: 0
other: 0
total: 2004335
transactions: 69115 (230.38 per sec.)
queries: 2004335 (6680.88 per sec.)
ignored errors: 0 (0.00 per sec.)
reconnects: 0 (0.00 per sec.)
General statistics:
total time: 300.0095s
total number of events: 69115
Latency (ms):
min: 5.66
avg: 8.68
max: 135.92
95th percentile: 10.65
sum: 599852.74
Threads fairness:
events (avg/stddev): 34557.5000/35.50
execution time (avg/stddev): 299.9264/0.00
Results ProxySQL connection to MySQL 8.0
SQL statistics:
queries performed:
read: 1930182
write: 0
other: 0
total: 1930182
transactions: 66558 (221.85 per sec.)
queries: 1930182 (6433.69 per sec.)
ignored errors: 0 (0.00 per sec.)
reconnects: 0 (0.00 per sec.)
General statistics:
total time: 300.0100s
total number of events: 66558
Latency (ms):
min: 5.85
avg: 9.01
max: 138.03
95th percentile: 11.65
sum: 599890.62
Threads fairness:
events (avg/stddev): 33279.0000/7.00
execution time (avg/stddev): 299.9453/0.00
Summary:
| type | transactions / sec | queries / sec | 95% latency (ms) |
|---|---|---|---|
| direct -> MySQL 5.7 | 396.77 | 11506.45 | 6.21 |
| direct -> MySQL 8.0 | 357.90 | 10379.20 | 6.79 |
| Haproxy -> MySQL 5.7 | 323.97 | 9395.09 | 7.43 |
| Haproxy -> MySQL 8.0 | 321.03 | 9309.79 | 7.43 |
| Traefik -> MySQL 5.7 | 327.21 | 9488.97 | 7.43 |
| Traefik -> MySQL 8.0 | 300.45 | 8712.97 | 7.84 |
| ProxySQL -> MySQL 5.7 | 230.38 | 6680.88 | 10.65 |
| ProxySQL -> MySQL 8.0 | 221.85 | 6433.69 | 11.65 |
We see a difference between MySQL 5.7 and 8.0 where we have about 90%
throughput with MySQL 8.0. Both TCP proxies give us about 80% throughput
compared to direct connection to MySQL. What is really surprising is that
ProxySQL only gives us a little below 60% throughput. We must also note that
the caching which is enabled in ProxySQL does not work because sysbench only
does prepared statements and those won’t be cached by ProxySQL as described in
the documentation of Limitiations of query cache.
Local vs Remote
When we look at our local results versus our remote results - and the test machines are in the same datacenter - the difference in throughput local vs remote is fairly high.
All results and some fun tests with long distance throughput and latency can be found in this calc sheet.
Conclusion
| type | transactions / sec | queries / sec | 95% latency (ms) | percentage |
|---|---|---|---|---|
| MySQL 5.7 | 1098.07 | 31844.15 | 2.03 | 100% |
| MySQL 8.0 | 933.37 | 27067.71 | 2.61 | 85% |
| direct -> 5.7 | 396.77 | 11506.45 | 6.21 | 36.1% |
| direct -> 8.0 | 357.90 | 10379.20 | 6.79 | 32.6% |
| Haproxy -> 5.7 | 323.97 | 9395.09 | 7.43 | 29.5% |
| Haproxy -> 8.0 | 321.03 | 9309.79 | 7.43 | 29.2% |
| Traefik -> 5.7 | 327.21 | 9488.97 | 7.43 | 29.8% |
| Traefik -> 8.0 | 300.45 | 8712.97 | 7.84 | 27.4% |
| ProxySQL -> 5.7 | 230.38 | 6680.88 | 10.65 | 21% |
| ProxySQL -> 8.0 | 221.85 | 6433.69 | 11.65 | 20.2% |
If we really want to have maximum throughput of our MySQL queries we need a local installation of MySQL since best case we already lose over 60% throughput once we introduce a real network connection. So this is very bad for applications that lean heavy on their MySQL backend.
Sadly enough we don’t get a proper view of what ProxySQL could potentially do in terms of caching. But if we look at that we serve mostly PHP applications and many of those use prepared statements everywhere it could be that ProxySQL can’t help us too much in that area. We’ll have to find out in a real life test (TODO).