Oracle published documentation about how to use development versions of the open source packages which are part of the Solaris 11.3 distribution:
IPS and Solaris userland sources have been moved from hg repos on hg.java.net to github:
Doing review of last two months zabbix DB activity I found that for less than week our zabbix DB backend has been doing regularly +2k selects/s reaching in peak more than 6.5k selects/s. It was due of one of our customers have been experimenting with mirroring some monitoring data over zabbix API. I’ve been carefully watching this and informing people around however all zabbix web frontend users not been able to notice selects rate increase as all additional selects have been hitting well cached DB data in ZFS ARC.
Here is the graph:
And last graph showing CPU usage in the same period on this machine:
As it is possible to see that impact on CPU usage effectively was zero/null 🙂
More than a month ago I made a Zabbix MySQL DB backend upgrade from 5.5 to 5.6. I made this upgrade to open up possibilities of tracing SQL traffic over USDT (User Statically Defined Tracing) probes and correlate those observations with other OS layer data. I’ve not been expecting some significant improvements as so far all what I’ve read about MySQL 5.6 improvements was mainly about slave synchronization speedup.
In below last 2 months graphs is possible to spot significant improvements on used storage side.
Below is IO graph of the SSD device which is used as L2ARC and ZIL device where is possible to spot significant reduction of write IOs.
Even bigger reduction of write IOs is possible to observe on single spindle disk IO graph.
Looks like with MySQL 5.5 we’ve been quite close to maximum IO bandwidth on single spindle. Upgrade to MySQL 5.6 caused huge reductions of write IOs which effectively is casing that it will be not necessary to upgrade hardware platform. In this case zabbix DB backend zpool uses 6 spindles working in RAID1+0 with one SSD used as zpool L2ARC+ZIL devices.
Before this upgrade I remember that on observing activity on DB backend using oneliner like “while true; do mysqladmin processlist | grep -v Sleep; sleep 1; do” usually was possible to see between 5 to 15 selects. After upgrade this list now no longer than 2-5 still running queries (mostly selects) so avg latency of all queries has been reduced. However I was not able easily observe SQL queries latency distribution as on 5.5 from Solaris 11.3 GA MySQL 5.5 was without USDT DTrace probes.
#mysql, #solaris, #zabbix