Re: Stalls on PGSemaphoreLock - Mailing list pgsql-performance

From Matheus de Oliveira
Subject Re: Stalls on PGSemaphoreLock
Date
Msg-id CAJghg4+gc6MEGwvsnJuKGbDSJ06kO=Vb4r70qi8WekqkgkT4Kw@mail.gmail.com
Whole thread Raw
In response to Re: Stalls on PGSemaphoreLock  (Matthew Spilich <mspilich@tripadvisor.com>)
List pgsql-performance

On Tue, Apr 22, 2014 at 12:12 PM, Matthew Spilich <mspilich@tripadvisor.com> wrote:
Hi all -  I am a little delayed in reporting back on this issue, but it was indeed the hugepage defrag setting that was the cause of my issue.

The transparent huge pages features seems so bogus for database workloads, that it is one of the first things I disable on new servers (I have tried to let it enabled sometimes, but every time the system was better with it disabled).
 
  One item that we noticed as we were testing this issue that I wanted to report back to the forum is that these settings
...
Were not sicky on reboot for my version of CentOS, which probably explains why I thought this was disabled already only to have it crop back up.   Anyway, I wanted to report back these findings to close the loop on this and to thank the community again for their support.

Just changing files at /sys/ is not permanent, so I recommend adding these commands into your /etc/rc.local file:

    test -f /sys/kernel/mm/transparent_hugepage/enabled && echo never > /sys/kernel/mm/transparent_hugepage/enabled
    test -f /sys/kernel/mm/transparent_hugepage/defrag && echo never > /sys/kernel/mm/transparent_hugepage/defrag


The test's are just to make sure the file does exists, as its location changes depending on the distro you are using and may also change on kernel upgrades.

It is also possible to add transparent_hugepage=never on grub.conf file, but I personally dislike this option.

Regards,
--
Matheus de Oliveira
Analista de Banco de Dados
Dextra Sistemas - MPS.Br nível F!
www.dextra.com.br/postgres

pgsql-performance by date:

Previous
From: Matthew Spilich
Date:
Subject: Re: Stalls on PGSemaphoreLock
Next
From: Jeff Janes
Date:
Subject: Re: tsearch2, large data and indexes