Running a PostgreSQL database and trying to understand the bottleneck in my
setup. I now know there is a fdsync being issued to a database file. This one
takes up to 10 seconds.
The problem here is that all the other processes are entering a wait state (this
is in postgres itself. The syncing process holds a latch in exclusive mode, that
is needed by all others) waiting for this flush to finish. All my 128+ processes
accumulate 10 seconds, throttling my throughput for a total of 1280 seconds.
This is significant!
There is definitely a relation with the database checkpoints that when happening
consequently make the long fdsync show up.
I would like to delve in a little deeper. It could be the sheer number of pages
that are being written due the fdsycn. Wonder if someone has a dtrace script
snippet I could use here to understand a little better what is going on under
the hood.
BTW running ZFS where my intent log is on a RAID array which has 1GB of write
caches.
Thanks,
Paul
--
This message posted from opensolaris.org
