Message-ID: <42822177.3010607@sgi.com>
Date: Wed, 11 May 2005 10:15:03 -0500
From: Eric Sandeen <sandeen@sgi.com>
User-Agent: Mozilla Thunderbird 1.0 (X11/20041206)
MIME-Version: 1.0
To: Bub Thomas <thomas.bub@thomson.net>
CC: linux-xfs@oss.sgi.com, Braehler Uwe <uwe.braehler@thomson.net>,
        Lindenkreuz Morris <morris.lindenkreuz@thomson.net>,
        Waldschmidt Stefan <stefan.waldschmidt@thomson.net>
Subject: Re: Every new file goes into a new ag
References: <B79FAF8BB536314E859EA1963CFFD22237B57F@wdtssmail01.eu.thmulti.com>
In-Reply-To: <B79FAF8BB536314E859EA1963CFFD22237B57F@wdtssmail01.eu.thmulti.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
Sender: linux-xfs-bounce@oss.sgi.com
Errors-to: linux-xfs-bounce@oss.sgi.com
Precedence: bulk
Content-Length: 1725
Lines: 38

If you have a "rotorstep" systune in your xfs codebase you might play 
with that; otherwise try making the size of your inode bigger (at mkfs 
time) so that you don't get into inode32 mode.  (or, if you're on a 
64-bit machine, mount with the "inode64" option to make 64-bit inodes).

for large filesystems, xfs must ensure that inode numbers don't go over 
32 bits.  To do that, inodes are allocated in the lower part of the 
fileysstem, and files are allocated round-robin through the AGs.

rotorstep changes that round-robin behavior to switch to a new AG every 
(X) new files, instead of every (1) new file.

If you can get out of inode32 mode by either mounting with 64-bit inodes 
on a 64-bit machine, or making the inode size larger (affects the inode 
numbering scheme) then files created in a single directory will 
generally be allocated in the same AG.

-Eric

Bub Thomas wrote:
> Hi there,
> I just got XFS running under RedHat EL 3.0.
> Since I need high bandwidth for media playout, where in our case each
> and every frame of a film image sequence is a new file, I need file
> sequences being physically close to each other.
> I'm struggling with the fact that consecutive files on a new filesystem
> don't get consecutive inode numbers. It seems that every new file ends
> up in a new ag.
> Thus the read performance for consecutive files is very poor 
> I tried it with two different kernels which are
> 2.4.21-27.0.2.EL.sgi9.i386 and 2.4.21-15.0.4.EL.sgi3.i38. with both the
> effect is the same.
> My filesystem is located on a JBOD array on 14 FC-2 disks with less then
> 2 TByte and the kernel config CONFIG_LBD not set.
> The volume on the JBOD I use is made with LVM.
> Any help welcome.
> Thomas Bub