Date: Fri, 12 Feb 2010 11:03:48 -0300
From: Alvaro Herrera <alvherre@commandprompt.com>
To: Kevin Grittner <Kevin.Grittner@wicourts.gov>
Cc: hannu@krosing.net, aidan@highrise.ca, jesper@krogh.cc,
	pgsql-performance@postgresql.org
Subject: Re: moving pg_xlog -- yeah, it's worth it!
Message-ID: <20100212140348.GA3737@alvh.no-ip.org>
References: <4B750193020000250002F22D@gw.wicourts.gov>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <4B750193020000250002F22D@gw.wicourts.gov>
User-Agent: Mutt/1.5.20 (2009-06-14)

Kevin Grittner wrote:
> Hannu Krosing  wrote:
>  
> > Can it be, that each request does at least 1 write op (update
> > session or log something) ?
>  
> Well, the web application connects through a login which only has
> SELECT rights; but, in discussing a previous issue we've pretty well
> established that it's not unusual for a read to force a dirty buffer
> to write to the OS.  Perhaps this is the issue here again.  Nothing
> is logged on the database server for every request.

I don't think it explains it, because dirty buffers are obviously
written to the data area, not pg_xlog.

> I wonder if it might also pay to make the background writer even more
> aggressive than we have, so that SELECT-only queries don't spend so
> much time writing pages.

That's worth trying.

> Anyway, given that these are replication
> targets, and aren't the "database of origin" for any data of their
> own, I guess there's no reason not to try asynchronous commit. 

Yeah; since the transactions only ever write commit records to WAL, it
wouldn't matter a bit that they are lost on crash.  And you should see
an improvement, because they wouldn't have to flush at all.

-- 
Alvaro Herrera                                http://www.CommandPrompt.com/
PostgreSQL Replication, Consulting, Custom Development, 24x7 support