DomainKey-Signature: a=rsa-sha1; c=nofws; d=gmail.com; s=beta;
	h=received:from:to:subject:date:user-agent:cc:references:in-reply-to:mime-version:content-type:content-transfer-encoding:content-disposition:message-id;
	b=Jhtzq9zs0tGzWZAbpHC0dW15Lmi2FpQSIiLkPe4hPwheuu7rWi5LvOUWT2G3uto8TlfcDGbUJOAeUeNqQp8tz9vCkDW30rv06zz48wfsZNRe8dLSpTxVGUXXtN/qQGEdhml4WojjNVVPlUGOX0y721c8nwZyolO+diO/4B2R/20=
From: Dimitri <dimitrik.fr@gmail.com>
To: pgsql-performance@postgresql.org
Subject: Re: Parallel Vacuum
Date: Thu, 22 Mar 2007 19:24:38 +0100
User-Agent: KMail/1.9.3
Cc: Michael Stone <mstone+postgres@mathom.us>
References: <200703221435.54408.dimitrik.fr@gmail.com>
	<200703221655.03068.dimitrik.fr@gmail.com>
	<20070322171007.GH11402@mathom.us>
In-Reply-To: <20070322171007.GH11402@mathom.us>
MIME-Version: 1.0
Content-Type: text/plain;
  charset="iso-8859-1"
Content-Transfer-Encoding: 7bit
Content-Disposition: inline
Message-Id: <200703221924.39278.dimitrik.fr@gmail.com>

Mike, 

you're right until you're using a single disk :) 
Now, imagine you have more disks - more I/O operations you may perform, and 
you'll need also a CPU time to process them :)  until you fully use one CPU 
per 'vacuumdb' - and then you stop... 

As well, even in case when CPU is not highly used by vacuumdb - single process 
is still not able to get a max performance of the storage array, just because 
you need several concurrent I/O running in the system to reach max 
throughput. And even filesystem might help you here - it's not all... More 
concurrent writers you have - higher performance you reach (until real 
limit)...

In my case I have a small storage array capable to give you more than 
500MB/sec and say 5000 op/s. All my data are striped throw all array disks. 
Single 'vacuumdb' process here become more CPU-bound rather I/O as it cannot 
fully load storage array... So, more vacuum processes I start in parallel - 
faster I'll finish database vacuuming.

Best regards!
-Dimitri


On Thursday 22 March 2007 18:10, Michael Stone wrote:
> On Thu, Mar 22, 2007 at 04:55:02PM +0100, Dimitri wrote:
> >In my case I have several CPU on the server and quite powerful storage box
> >which is not really busy with a single vacuum. So, my idea is quite simple
> > - speed-up vacuum with parallel execution (just an algorithm):
>
> Vacuum is I/O intensive, not CPU intensive. Running more of them will
> probably make things slower rather than faster, unless each thing you're
> vacuuming has its own (separate) disks. The fact that your CPU isn't
> pegged while vacuuming suggests that your disk is already your
> bottleneck--and doing multiple sequential scans on the same disk will
> definitely be slower than doing one.
>
> Mike Stone
>
> ---------------------------(end of broadcast)---------------------------
> TIP 3: Have you checked our extensive FAQ?
>
>                http://www.postgresql.org/docs/faq