Gluster users - Mar 2012 - adding bricks of diferent size to gluster

Dear all,

I've been trying to locate documentation about gluster functionality about
using bricks of diferent sizes and did not find anything useful.

What I was trying to answer was:

*How does using bricks of different size in a distributed replicated volume
affects performance?

I think I read somewhere that it was highly advisable to use bricks of the
same size but I can not locate where I read such statement? Is it true?

Case it's depending on the version, I'm using mostly 3.2.4 version.

Thank in advance,
Samuel.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://supercolony.gluster.org/pipermail/gluster-users/attachments/20120305/bccf60fa/attachment.html>

On 03/05/2012 09:39 AM, gluster-users-request at gluster.org
wrote:
> Date: Mon, 5 Mar 2012 10:29:22 +0100
> From: samuel<samu60 at gmail.com>
> Subject: [Gluster-users] adding bricks of diferent size to gluster
> To:gluster-users at gluster.org
> Message-ID:
> 	<CAOg=WDch3b72zN54B-11jQ_FPbF6MDe8qV86p1KfBG4jY76Mjg at
mail.gmail.com>
> Content-Type: text/plain; charset="iso-8859-1"
>
> Dear all,
>
> I''ve been trying to locate documentation about gluster
functionality about
> using bricks of diferent sizes and did not find anything useful.
>
> What I was trying to answer was:
>
> *How does using bricks of different size in a distributed replicated volume
> affects performance?
>
> I think I read somewhere that it was highly advisable to use bricks of the
> same size but I can not locate where I read such statement? Is it true?
>
> Case it''s depending on the version, I''m using mostly
3.2.4 version.
>
> Thank in advance,
> Samuel.
Samuel,
The mechanism described in the following gluster-users posting describes 
how non uniform brick sizes are managed by GlusterFS.

http://gluster.org/pipermail/gluster-users/2011-September/008754.html

I don''t know if this approach carries a performance penalty, but I 
decided to move away from non uniform brick sizes because this 
configuration is not rigorously tested by the developers, as explained 
in the gluster-users posting below.

http://gluster.org/pipermail/gluster-users/2011-May/007788.html

The posting below describes how I have set up uniform brick sizes 
despite having different sized servers in my cluster.

http://gluster.org/pipermail/gluster-users/2011-August/008597.html

If you are considering using non uniform brick sizes please make sure 
that you set the min-free-disk parameter to an appropriate value in 
gigabytes.  I originally had non uniform brick sizes in my cluster and I 
had a lot of problems with small bricks filling up before larger bricks, 
resulting in "device full" errors even when there was plenty of space 
left in the volume.  The solution I eventually found is described in the 
following gluster-users posting.

http://gluster.org/pipermail/gluster-users/2011-November/009176.html

The "device full" problem was caused by the default value of 
min-free-disk being defined as a percentage rather than as a value in 
gigabytes.  To set the min-free-disk parameter to a value in GB follow 
the instructions described in the following gluster-users postings.

http://gluster.org/pipermail/gluster-users/2011-May/007794.html
http://gluster.org/pipermail/gluster-users/2011-August/008579.html

It is important to make sure the value of min-free-disk is big enough.  
I found that setting the value to 10% of the size of the smallest brick 
seemed to work, but when I set it to only ~1% of the brick size I still 
had problems with "device full" errors.


I hope this helps.
-Dan.