[ale] XFS multi-mount

[Greg Freemyer]

Attriel,

I assume you know that what you are doing is simply not supported by
most filesystems, right?

The "correct" thing to do would be to try a cluster filesystem. 

For what you describe, I think GFS from Redhat or OCFS from Oracle
would be the best GPL choices.  Both are designed  to have multiple
nodes accessing a common shared storage array.

Unfortunately neither of these is likely to be as fast as XFS.  

I would check out OCFS first.  It has some short-comings (no locking,
no acls, etc.),  but I don't think you need those features.

If you need XFS, you could also try mounting with the sync flag.  That
tells xfs to keep the servers write buffer flushed to disk at all
times.  Unfortunately, the sync flag really slows down XFS's
performance.

That would fix your write caching issues.

As Dow said, it is your read cache that is likely the real difficulty.
 Maybe you could post on the XFS list and someone could tell you of
way to invalidate all the read cache pages, thus forcing a physical
read.  If that could be done every 30 seconds or so, it might work for
you.

Greg
-- 
Greg Freemyer

On Mon, 20 Dec 2004 16:11:56 -0500 (EST), Dow_Hurst
<dhurst at mindspring.com> wrote:
> XFS is big on caching.  Have your tried a sync on the readonly machines?  You might need to look into fam more closely.  I'm guessing a bit but your xfsd is probably not checking on files it knows that it hasn't changed.  NFS with locking would circumvent that.  File Alteration Monitor, fam, might help in this situation.  Hope this helps and doesn't lead you down the wrong path.  I remember that SGI's implementation of XFS had some RPC based daemon for talking XFS stuff, but I never looked into what it was for.  It may be applicable and duplicated in the Open Source XFS.  You could look into that.
> Dow
> 
> -----Original Message-----
> From: attriel <attriel at d20boards.net>
> Sent: Dec 20, 2004 12:02 PM
> To: ale at ale.org
> Subject: [ale] XFS multi-mount
> 
> OK, so I've got a RAID device that is hosting all the images for an image
> repository site I manage.
> 
> I set it up with XFS b/c I didn't even want to think about the fsck on a
> pair of 1T partitions !!!
> 
> The plan was to hook the XFS up RW on one machine (admin) and then RO on
> the others.  That part works OK.  My problem comes in with when the admin
> server makes an update/adds new files/etc ... The RO systems don't see it.
> 
> In fact, I can't seem to get them to see it unless I umount the XFS from
> ALL servers, so that it is completely unmounted, then remount it.  Is
> there an option or setting I'm missing?  Some tool I can run that says "Go
> forth, and re-read the XFS data from the drive. Honest, it's more
> relevant" ?
> 
> I'd rather not add the overhead of an NFS solution, and my boss won't be
> particularly happy to hear about the forced outage of the system after
> pushing so much for a 24/7 reliable always available system design from me
> :/
> 
> --attriel
> _______________________________________________
> Ale mailing list
> Ale at ale.org
> http://www.ale.org/mailman/listinfo/ale
> 
> No sig.
> _______________________________________________
> Ale mailing list
> Ale at ale.org
> http://www.ale.org/mailman/listinfo/ale
>