Message-ID: <50B91012.7020202@kickstone.com>
Date: Fri, 30 Nov 2012 19:59:14 +0000
From: John Lister <john.lister@kickstone.com>
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64;
	rv:17.0) Gecko/17.0 Thunderbird/17.0
MIME-Version: 1.0
To: Merlin Moncure <mmoncure@gmail.com>
CC: pgsql-performance@postgresql.org
Subject: Re: Comparative tps question
References: <50B659EC.4080509@kickstone.com>
	<CAHyXU0xasEPfQ7d8N=0zH5i_N+Giid8x2rUdNHzabSqu8mBwYA@mail.gmail.com>
	<50B793C5.3070806@kickstone.com>
	<CAHyXU0zqN=xaRLM_uKZAfKWFGnSK6ROKfPMgE0rhmQCORoOD2A@mail.gmail.com>
In-Reply-To: 
 <CAHyXU0zqN=xaRLM_uKZAfKWFGnSK6ROKfPMgE0rhmQCORoOD2A@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 29/11/2012 17:33, Merlin Moncure wrote:
> Since we have some idle cpu% here we can probably eliminate pgbench as 
> a bottleneck by messing around with the -j switch. another thing we 
> want to test is the "-N" switch -- this doesn't update the tellers and 
> branches table which in high concurrency situations can bind you from 
> locking perspective.
Using -N gives around a 15% increase in tps with no major changes in 
load, etc. using more threads slightly drops the performance (as 
expected with only 32 "cores"). dropping it does give a slight increase 
(presumably because half the cores aren't real).

> one thing that immediately jumps out here is that your wal volume
> could be holding you up.  so it's possible we may want to move wal to
> the ssd volume.  if you can scrounge up a 9.2 pgbench, we can gather
> more evidence for that by running pgbench with the
> "--unlogged-tables" option, which creates the tables unlogged so that
> they are not wal logged (for the record, this causes tables to be
> truncated when not shut down in clean state).
I did notice that using -S drives the tps up to near 30K tps, so it is 
possibly the wal volume, although saying that I did move the pg_xlog 
directory onto the ssd array before posting to the list and the 
difference wasn't significant. I'll try and repeat that when I get some 
more downtime (I'm having to run the current tests while the db is live, 
but under light load).

I'll have a look at using the 9.2 pgbench and see what happens.
> yeah -- this will tell us raw seek performance of ssd volume which
> presumably will be stupendous.  2x is minimum btw 10x would be more
> appropriate.
>
> since you're building a beast, other settings to explore are numa
> (http://frosty-postgres.blogspot.com/2012/08/postgresql-numa-and-zone-reclaim-mode.html)
> and dell memory bios settings that are occasionally set from the
> factory badly (see here:
> http://bleything.net/articles/postgresql-benchmarking-memory.html).
Cheers for the links, I'd already looked at the numa stuff and disabled 
zone reclaim. I was looking at using the patch previously posted that 
used shared mode for the master process and then local only for the 
workers - excuse the terminology - but time constraints prevented that.
Made sure the box was in performance mode in the bios, unfortunately I 
spotted bens blog when I was setting the box up, but didn't have time to 
go through all the tests. At the time performance seemed ok (well better 
than the previous box :) - but having it live for a while made me think 
I or it could be doing better.

Anyway, I still think it would be nice to post tps results for compative 
purposes, so if I get a minute or two I'll create a site and stick mine 
on there.

John


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