Merlin Moncure <mmoncure@gmail.com> wrote:
> we have to be careful to rule out some underlying possible
> contributing factors before switching up things up to much.
Agreed.
> THP compaction in particular has plaguing servers throughout the
> company I work for;
I've seen many support tickets where turning off Transparent Huge
Page support has been the solution, but in all cases that I've seen
it is *system* CPU time that spikes when that is the problem, and
the OP here showed a graph where it was *user* CPU time spiking.
With high concurrency that is usually (in my experience) spinlocks
inside of PostgreSQL -- often spinlocks guarding transitions of hot
lightweight locks.
> we haven't figured out OP's system went loaded all of a sudden.
Agreed, although I feel that THP problems are not indicated because
system CPU time wasn't pegged, and write gluts seem unlikely based
on the IO wait numbers.
--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company