> Are you able to take some 'perf top' during high CPU spike and see
> what's burning CPU there? Though the issue is related to blocking, but
> high CPU spikes may hint some spinning to acquire behavior.
Will do, although hopefully the spikes were only growing pains after the
upgrade.
> If your previous relation size is smaller than after upgrade, that's a
> signal that you do have holes in relation, thus extension can be
> avoided sometimes for new tuples.
The relation between high CPU and page splits is not immediately obvious
to me.
We run with synchronous_commit off, but there does seem to be a peak in
I/O requests around the CPU spikes.
Is a page split by nature a synchronous I/O activity? And do the other
connections wait in some kind of CPU intensive form (like a spinlock?)
Kind regards,
Andomar