Re: Several optimization options (config/hardware) - Mailing list pgsql-performance

From Martin Grotzke
Subject Re: Several optimization options (config/hardware)
Date
Msg-id 4FA797C1.7040603@googlemail.com
Whole thread Raw
In response to Re: Several optimization options (config/hardware)  ("Albe Laurenz" <laurenz.albe@wien.gv.at>)
List pgsql-performance
On 05/04/2012 09:57 AM, Albe Laurenz wrote:
> Martin Grotzke wrote:
>>> You could try different kernel I/O elevators and see if that
>>> improves something.
>>>
>>> I have made good experiences with elevator=deadline and
>>> elevator=noop.
>
>> Ok, great info.
>>
>> I'm not sure at which device to look honestly to check the current
>> configuration.
>>
>> mount/fstab shows the device /dev/mapper/VG01-www for the relevant
>> partition. When I check iostat high utilization is reported for
>> the devices dm-4 and sda (showing nearly the same numbers for util
>> always),
>> so I suspect that dm-4 is mapped on sda.
>
> Use the option -N of "iostat" to see long device names. You can use
> "lvm" to figure out the mapping.
iostat with -N shows VG01-www for dm-4. For lvm/lvdisplay/dmsetup I get
"Permission denied" as I have no root/sudo permissions. I need to check
this with our hosting provider (hopefully we have a call today).


>> This is the current config: $ cat /sys/block/sda/queue/scheduler
>> noop anticipatory deadline [cfq] $ cat
>> /sys/block/dm-4/queue/scheduler none
>
> Do you mean literal "none" or do you mean that the file is empty?
"none" was the output of `cat /sys/block/dm-4/queue/scheduler`.


>> Which of them should be changed? I'll discuss this also with our
>> hosting provider next week, he'll know what has to be done.
>
> I'd just add "elevator=deadline" to the kernel line in
> /etc/grub.conf and reboot.  At least if it is a dedicated database
> machine.
>
> But of course you want to change it on the fly first to test - not
> knowing the answer to your question, I would change it in both
> devices if I can.
Ok, makes sense.

Cheers,
Martin


Attachment

pgsql-performance by date:

Previous
From: Richard Jones
Date:
Subject: Re: Partitioned/inherited tables with check constraints causing slower query plans
Next
From: Robert Klemme
Date:
Subject: Re: Result Set over Network Question