A benchmark should be repeatable and determinisitic
sysbench --test=fileio --fiile-totoal-size=32G --file-block-size=16K
--file-extra-flags=direct
--rant-init-on
--file-io-mode
--fileitest-mode
Drop system caches
echo 3 > /proc/sys/vm/dropcaches
jitter plot
dots become larger when multiple hits at the same point
Linux native AIO, Async I/O
for background tasks
Benchmark mixed I/O
Replication is single threader
create database sysbend
grean all on sbtest.*
to sbtest@localhost identifed by sbtest
-oltp-table-size
-iltp-table count
--mysql-user=
--mysql-password=
--mysql-host=
-report-interval
Reconnecting
db_disconnect
emulates connection churning
netstat -teanp
Limit number of time_wait connections
echo 1000 > /proc/sys/net/ipv4/tcp_max_tw_buckets
time_wait connections are for clients to reconnect.
(waiting for FIN packet)