I have been facing problem with my database server quite a month, Below are the observations that I see when it hits the top.
- load average 40 to 50
- CPU % - 400%
- idle % - 45%
- wait % - 11%
- vmstat procs r-> 14 and b-> 5
And then drains down within 5 minutes. And when I check the show processlist I see queries for DML and SQL are halted for some minutes. And it processes very slowly. Whereas each query are indexed appropriately and there will be no delay most of the time it returns less than 1 second for any query that are being executed to server the application.
Mysql Version : 5.0.77
OS : CentOS 5.4
Mem: 16GB RAM (80% allocated to INNODB_BUFFER_POOL_SIZE)
Database Size: 450 GB
16 Processor & 4 cores
Not in per-table model.
TPS ranges 50 to 200.
Master to a Slave of the same configuration and seconds behind is 0.
Below url shows show innodb status \G and show open tables; at the time spike. And this reduced within 5 minutes. Sometimes rare scenarios like once in two months I see the processes takes more than 5 to 8 hours to drain normal. All time I notice the load processor utilization and how it gradually splits its task and keep monitoring the process and innodb status and IO status. I need not do anything to bring it down. It servers the applications promptly and after some time it drains down to normal. Can you find anything suspicious in the url if any locks or OS waits any suggestion to initially triage with or what could have caused such spikes ?
http://tinyurl.com/bm5v4pl -> "show innodb status \G and show open tables at DB spikes."
Also there are some concerns that I would like to share with you.
Recently I have seen a table that gets inserts only about 60 per second. It predominantly locks for a while waiting for auto-inc to get released. And thus subsequent inserts stays in the processlist tray. After a while the table gets IN_USE of about 30 threads and later I don't know what it makes to free them and clears the tray. (At this duration the load goes more than 15 for 5 minutes)
Suppose if you say application functionality should be shapped to best suite the DB server to react. There are 3 to 5 functionalities each are independent entities in schema wise. Whenever I see the locks it gets affected to all other schemas too.
Now what makes best fuzzy is the last one. I see slave keeps in synch with master with a delay of 0 second all time whereas slave has a single thread SQL operation that is passed from relay IO that which acts in FIFO model from the binary logs where Master had generated. When this single headed slave can keep the load less and have the operations upto-date, should the concurrent hits are really made to be concurrent for the functionalities which I assume making the possible IO locks in OS level. Can this be organized in application itself and keep the concurrent tenure density thinner?