On 4/19/15 3:08 PM, Jim Nasby wrote:
>> > I do suspect your pgfree/s is very high though; putting 200k
>> pages/s on
>> > the free list seems like something's broken.
>> >
>> The system has constant and considerable load of small writes. The
>> pg_activity tool shows 300 IOPs sustained (it claims max IPs above
>> 11000.) Postgres 9.3 had a comparable pgfree/s.
>
> That leads me to a new theory... you may be running into problems
> finding free buffers in the buffer pool. We need to have a buffer before
> we can extend a relation, and if you have a lot of pressure on shared
> buffers it can take quite a bit of CPU to find one. To make matters
> worse, that search for a buffer takes place while holding the extension
> lock.
>
> Would you be able to get a stack trace of a backend that's holding an
> extension lock? Or maybe perf would provide some insight.
BTW,
http://postgresql.org/message-id/flat/20150329185619.GA29062@alap3.anarazel.de
has some useful info about this.
--
Jim Nasby, Data Architect, Blue Treble Consulting
Data in Trouble? Get it in Treble! http://BlueTreble.com