On 4/19/15 4:24 AM, Andomar wrote:
To put the top question first:
How can table extension locks explain a a massive spike in CPU usage?
I can imagine 400 connections waiting on disk I/O, but then, wouldn't
they all be sleeping?
Not necessarily. Spinlocks don't put the process to sleep, but they're
also supposed to be very short lived.
> Ok, that's a MAJOR hint, because relation 1249 is a system catalog;
> namely pg_attribute. So I think what's happening here is that your
> catalog has become horrifically bloated. I'm 99% certain that VACUUM ALL
> will not vacuum the catalog tables.
>
> Do you by chance have autovacuum turned off?
>
> A manual VACUUM VERBOSE pg_attribute might provide some immediate
relief.
>
Autovacuum is turned on. In addition, we do a manual VACUUM ALL at
night. VACUUM VERBOSE pg_attribute ran in 0 seconds and processed a few
hundred rows.
> Are you using a connection pool? Establishing 50 new database
> connections per second won't do anything to help performance...
>
As I understand it, a pool reduces network and CPU load. We have never
seen any issues with those. So the extra monitoring and maintenance
cost of a pool seems hard to justify.
Well, it sounds like you are CPU bound here... :P I don't know if this
is related or not, but it wouldn't hurt. If you install pg_bouncer on
the database server itself (which it's designed for) it shouldn't add
much maintenance cost.
> I think what that means is that there was suddenly a big spike in memory
> demand at the OS level, so now the OS is frantically dumping cached
> pages. That in itself won't explain this, but it may be a clue.
>
We monitor memory usage with Cacti. It's a dedicated server and nearly
all memory is used as cache. If a script runs and demands memory, that
becomes visible as cache is cleared out. There is no change in the
amount of memory used as cache around the outage.
> In order to extend a relation we need to ask the filesystem to actually
> extend the file (which presumably means at least writing some metadata
> to disk), and then I think we create a WAL record. Creating the WAL
> record won't by itself write to disk... *unless* the wal_buffers are all
> already full.
>
I have a question here, we have "synchronous_commit = off". So when
Postgres extends a page, would it do that just in memory, or does part
of the extend operation require synchronous I/O?
Turning that off doesn't mean there will never be an fsync, it just
means that we don't wait for one before returning from COMMIT. I don't
think relation extension itself requires a fsnyc, but see below.
> So if you also see an I/O spike when this happens you could well
> just be starved from the I/O system (though obviously it'd be
> better if we handled that situation more elegantly than this).
The SAR data shows no increase in pgpgin/s and pgpgout/s, which if I
understand it correctly, means that there is no I/O spike. There is
however an enormous increase in CPU usage.
I'm not familiar enough with SAR to know if that's correct or not;
iostat would be a good way to confirm it.
> 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.
Would you know a good resource to get more knowledgeable about pgfree,
pgpin, pgsteal?
Unfortunately I'm not strong on the system tools.
--
Jim Nasby, Data Architect, Blue Treble Consulting
Data in Trouble? Get it in Treble! http://BlueTreble.com
--
Sent via pgsql-general mailing list (pgsql-general@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-general
