Обсуждение: Sudden crazy high CPU usage

I'm running postgresql 9.3 on a production server. An hour ago, out of the "blue", I ran into an issue I have never
encounteredbefore: my server started to use CPU as crazy. The server is a standard ubuntu 12.04 LTE installation
runningonly Postgres and Redis. 

The incident can be seen on the in numbers below:

https://s3-eu-west-1.amazonaws.com/autouncle-public/other/cpu.png

I imidiatly took a look at pg_stat_activity but nothing in there seemed suspicious. I also had a look at the postgres
log,but nothing was in there too. I have pg_stat_statements running, so I reseted that one, and nothing really
suspiciousoccurred in there, expect for the fact, that all queries were taking 100x times longer than usual. 

I have tried the following with no luck:

    • Restart clients connecting to the db
    • Restart postgres
    • Restart the whole server

I have run memory tests on the server as well, and nothing seems to be wrong.

No changes in any software running on the servers has been made within the last 24 hours.

The question is: I have a streaming replication server running, which I have now done a failover to, and it runs fine.
HoweverI still have no clue why my master suddenly has become so CPU consuming, and how I can debug / trace it further
down?

On Mon, Mar 31, 2014 at 5:25 AM, Niels Kristian Schjødt
<nielskristian@autouncle.com> wrote:
> I'm running postgresql 9.3 on a production server. An hour ago, out of the "blue", I ran into an issue I have never
encounteredbefore: my server started to use CPU as crazy. The server is a standard ubuntu 12.04 LTE installation
runningonly Postgres and Redis. 
>
> The incident can be seen on the in numbers below:
>
> https://s3-eu-west-1.amazonaws.com/autouncle-public/other/cpu.png
>
> I imidiatly took a look at pg_stat_activity but nothing in there seemed suspicious. I also had a look at the postgres
log,but nothing was in there too. I have pg_stat_statements running, so I reseted that one, and nothing really
suspiciousoccurred in there, expect for the fact, that all queries were taking 100x times longer than usual. 
>
> I have tried the following with no luck:
>
>         * Restart clients connecting to the db
>         * Restart postgres
>         * Restart the whole server
>
> I have run memory tests on the server as well, and nothing seems to be wrong.
>
> No changes in any software running on the servers has been made within the last 24 hours.
>
> The question is: I have a streaming replication server running, which I have now done a failover to, and it runs
fine.However I still have no clue why my master suddenly has become so CPU consuming, and how I can debug / trace it
furtherdown? 

Using linux 6? One possible culprit is "Transparent Huge Page
Compaction".  It tends to hit severs with a lot of memory, especially
if they've configured a lot of shared buffers.  Google it a for a lot
of info.

There may be other issues masquerading as this one but it's the first
thing to rule out.  Symptoms are very high cpu utilization and poor
performance that strikes without warning and then resolves also
without warning (typically seconds or minutes after the event).

For starters, take a look at the value of:

/sys/kernel/mm/redhat_transparent_hugepage/enabled

And do some due diligence research.

merlin

On Mon, Mar 31, 2014 at 8:24 AM, Niels Kristian Schjødt
<nielskristian@autouncle.com> wrote:
>
> Thanks, this seems to persist after a reboot of the server though, and I have never in my server's 3 months life time
experiencedanything like it. 

Could it be overheating and therefore throttling the cores?

Also another thing to look at on large memory machines with > 1 CPU
socket is zone_reclaim_mode being set to 1. Always set it to 0 on a
linux machine running postgres.

Thanks, I don’t think overheating is an issue, it’s a large dell server, and I have checked the historic CPU
temperaturein the servers control panel, and no overheating has shown. 

Zone_reclaim_mode is already set to 0

Den 31/03/2014 kl. 16.50 skrev Scott Marlowe <scott.marlowe@gmail.com>:

> On Mon, Mar 31, 2014 at 8:24 AM, Niels Kristian Schjødt
> <nielskristian@autouncle.com> wrote:
>>
>> Thanks, this seems to persist after a reboot of the server though, and I have never in my server's 3 months life
timeexperienced anything like it. 
>
> Could it be overheating and therefore throttling the cores?
>
> Also another thing to look at on large memory machines with > 1 CPU
> socket is zone_reclaim_mode being set to 1. Always set it to 0 on a
> linux machine running postgres.

On Mon, Mar 31, 2014 at 3:25 AM, Niels Kristian Schjødt
<nielskristian@autouncle.com> wrote:
> I'm running postgresql 9.3 on a production server. An hour ago, out of the "blue", I ran into an issue I have never
encounteredbefore: my server started to use CPU as crazy. The server is a standard ubuntu 12.04 LTE installation
runningonly Postgres and Redis. 
>
> The incident can be seen on the in numbers below:
>
> https://s3-eu-west-1.amazonaws.com/autouncle-public/other/cpu.png

The increase doesn't look so sudden. My guess is that the server got
some new activity. The advice is to setup the statistics collecting
script by the link [1] and review the results for a period of hour or
so. It shows charts of statements by CPU/IO/calls with aggregated
stats, so you could probably find out more than with pure
pg_stat_statements.

[1] https://github.com/grayhemp/pgcookbook/blob/master/statement_statistics_collecting_and_reporting.md

--
Kind regards,
Sergey Konoplev
PostgreSQL Consultant and DBA

http://www.linkedin.com/in/grayhemp
+1 (415) 867-9984, +7 (901) 903-0499, +7 (988) 888-1979
gray.ru@gmail.com

On 2014-03-31 19:16:58 +0200, Niels Kristian Schjødt wrote:
> Yes, I could install “perf”, though I’m not familiar with it. What would i do? :-)

As root:
perf record -a sleep 5
perf report > my-nice-perf-report.txt

And then send the my-nice-perf-report.txt file.

Locally it's much nicer to see the output using "perf report" without
redirect into a file, you'll get an interactive UI.

Greetings,

Andres Freund

--
 Andres Freund                       http://www.2ndQuadrant.com/
 PostgreSQL Development, 24x7 Support, Training & Services

On Wed, Apr 2, 2014 at 7:16 AM, Andres Freund <andres@2ndquadrant.com> wrote:
> On 2014-03-31 19:16:58 +0200, Niels Kristian Schjødt wrote:
>> Yes, I could install "perf", though I'm not familiar with it. What would i do? :-)
>
> As root:
> perf record -a sleep 5
> perf report > my-nice-perf-report.txt
>
> And then send the my-nice-perf-report.txt file.
>
> Locally it's much nicer to see the output using "perf report" without
> redirect into a file, you'll get an interactive UI.
The Postgres wiki has a page dedicated to perf as well here:
https://wiki.postgresql.org/wiki/Profiling_with_perf

Regards,
--
Michael