Message-ID: <52A0F79C.3060200@catalyst.net.nz>
Date: Fri, 06 Dec 2013 11:01:00 +1300
From: Mark Kirkwood <mark.kirkwood@catalyst.net.nz>
User-Agent: Mozilla/5.0 (X11; Linux x86_64;
 rv:24.0) Gecko/20100101 Thunderbird/24.1.1
MIME-Version: 1.0
To: Skarsol <skarsol@gmail.com>, Scott Marlowe <scott.marlowe@gmail.com>
CC: desmodemone <desmodemone@gmail.com>,
 "pgsql-performance@postgresql.org" <pgsql-performance@postgresql.org>
Subject: Re: WAL + SSD = slow inserts?
References: 
 <CAMt8e=Ee9BRLvNk3qq3vw-KKu-myMOpbx+9y4ABQee1dLcTq7Q@mail.gmail.com>
	<CAEs9oFk4bFCA7cS=Q-EkN6joMbqwrGGpL8v8dU8d9S1rpA0HGA@mail.gmail.com>
	<CAMt8e=FxA=0x8G-BTuV00BNGSz_-zkykrWG465mg0LBmO+sRxA@mail.gmail.com>
	<CAOR=d=2zC19+SJLJ2SBNOb67imXGC8qDZEy_NzVOygeMUVOXUQ@mail.gmail.com>
 <CAMt8e=GbJw3yWK9ykusv3OM8wPS0k4ciOCOi28Y9NABmP0a9KA@mail.gmail.com>
In-Reply-To: 
 <CAMt8e=GbJw3yWK9ykusv3OM8wPS0k4ciOCOi28Y9NABmP0a9KA@mail.gmail.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk
Sender: pgsql-performance-owner@postgresql.org

On 06/12/13 05:13, Skarsol wrote:
> On Thu, Dec 5, 2013 at 9:50 AM, Scott Marlowe <scott.marlowe@gmail.com>wrote:
>
>> On Thu, Dec 5, 2013 at 8:16 AM, Skarsol <skarsol@gmail.com> wrote:
>>> psql (PostgreSQL) 9.2.5
>>> Red Hat Enterprise Linux Server release 6.4 (Santiago)
>>> Linux 2.6.32-358.6.1.el6.x86_64 #1 SMP Fri Mar 29 16:51:51 EDT 2013
>> x86_64
>>> x86_64 x86_64 GNU/Linux
>>> All relevant filesystems are ext4
>>>
>>> Changes from defaults:
>>> max_connections = 500
>>> shared_buffers = 32000MB
>>> temp_buffers = 24MB
>>> work_mem = 1GB
>>> maintenance_work_mem = 5GB
>>> wal_level = archive
>>> wal_buffers = 16MB
>>> checkpoint_completion_target = 0.9
>>> archive_mode = on
>>> archive_command = 'test ! -f /databases/pg_archive/db/%f && cp %p
>>> /databases/pg_archive/db/%f'
>>> effective_cache_size = 64000MB
>>> default_statistics_target = 5000
>>> log_checkpoints = on
>>> stats_temp_directory = '/tmp/pgstat'
>> OK I'd make the following changes.
>> 1: Drop shared_buffers to something like 1000MB
>> 2: drop work_mem to 16MB or so. 1GB is pathological, as it can make
>> the machine run out of memory quite fast.
>> 3: drop max_connections to 100 or so. if you really need 500 conns,
>> then work_mem of 1G is that much worse.
>>
>> Next, move pg_xlog OFF the SSDs and back onto spinning media and put
>> your data/base dir on the SSDs.
>>
>> SSDs aren't much faster, if at all, for pg_xlog, but are much much
>> faster for data/base files.
>>
>> Also changing the io schduler for the SSDs to noop:
>>
>>
>> http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/tree/Documentation/block/switching-sched.txt?id=HEAD
>>
> Changing the scheduler to noop seems to have had a decent effect. I've made
> the other recommended changes other than the connections as we do need that
> many currently. We're looking to implement pg_bouncer which should help
> with that.
>


What model SSD are you using? Some can work better with deadline than 
noop (as their own scheduling firmware may be pretty poor). Also, check 
if there are updates for the SSD firmware. I have a couple of Crucial 
M4s that changed from being fairly average to very fast indeed after 
getting later firmware...

Cheers

Mark


-- 
Sent via pgsql-performance mailing list (pgsql-performance@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-performance