[Top] [All Lists]

Re: Maximum log stripe size

To: Stéphane Doyon <sdoyon@xxxxxxxxx>
Subject: Re: Maximum log stripe size
From: Steve Lord <lord@xxxxxxx>
Date: 25 Mar 2003 13:11:29 -0600
Cc: Nathan Scott <nathans@xxxxxxx>, linux-xfs@xxxxxxxxxxx
In-reply-to: <Pine.LNX.4.44.0303251247020.5702-100000@xxxxxxxxxxxxxxxxxx>
References: <Pine.LNX.4.44.0303251247020.5702-100000@xxxxxxxxxxxxxxxxxx>
Sender: linux-xfs-bounce@xxxxxxxxxxx
On Tue, 2003-03-25 at 11:50, Stéphane Doyon wrote:
> Hi,
> mkfs.xfs insists that the log stripe unit (-l sunit) must be
> smaller than 256K.
> See xfs_mkfs.c around line 1730:
>         if ((lsunit * blocksize) > 256 * 1024) {
>                 fprintf(stderr,
> _("log stripe unit (%d bytes) is too large for kernel to handle (max 
> 256k)\n"),
>                         (lsunit * blocksize));
>                 exit(1);
>         }
> Can anyone please tell me whether this is really necessary?
> Our SCSI RAID controllers would much prefer stripes in the order of 1MB.
> And the kernel doesn't seem to mind.

You really really do not want to make them this big. The reason for
the maximum size is the maximum size of iclog buffers - which they
are written from. These are capped at 256K it is very hard to guarantee
allocations of large chunks of kernel memory.

Secondly, the larger the stripe, the more log space is burned in 
padding. The more log space we use, the faster metadata has to be
flushed to disk to allow old log space to be reused. So a large
stripe size is probably going to mean more I/O for the same number
of filesystem operations.



Steve Lord                                      voice: +1-651-683-3511
Principal Engineer, Filesystem Software         email: lord@xxxxxxx

<Prev in Thread] Current Thread [Next in Thread>