[Linux-cluster] Mirroring GFS

[Daniel Phillips]

On Thursday 28 April 2005 10:42, Andrea Manca wrote:
> I've a couple of question about interaction between MD devices
> (software RAID), GNBD, and GFS.
>
> I'm testing a cluster with four nodes: two gfs and two gnbd nodes.
> GNBD1 and GNBD2 export each one a partition over the LAN;
> GFS1 and GFS2 import that exported block devices.
>
> I'm interested in mirroring the information in GNBD1 over GNBD2, so
> i've thought about creating a md device composed by
>
> /dev/gnbd/GNBD1_PRIMARY
> /dev/gnbd/GNBD2_MIRROR
>
> in RAID1 configuration and gfs formatted.
> It seem's a reasonable configuration for me, but in:
>
> https://www.redhat.com/archives/linux-cluster/2004-October/msg00055.h
>tml
>
> a similar solution (except for a RAID5 configuration), is discussed
> and marked as impossible, because Software Raid in linux isn't
> cluster aware.

This is the block driver you are looking for:

   http://sourceware.org/cluster/ddraid/

> This could lead an inconsistency in parity blocks 
> which aren't locked correctly.
> In my opinion two mirrored disks doesn't need a block level locking,
> because parity block aren't present.
>
> i've tried this on 4 VMWare machines and it seem work well if you
> write (or read) something from GFSx nodes.
>
> Could my solution get working or there is a similar issue applicable
> in my situation?
> In this second case, are any workarounds possible?
>
>
> I really appreciate any comment or explanation.

Straight-up md raid5 does read-before-write with no global 
synchronization.  GFS does not know about this and can't know about it.   
Neither does md.  If it works for you, you are just lucky - this time.

Besides that problem, there is the matter of synchronizing resync IO 
across the cluster and interfacing to the cluster membership etc.  
DDRaid addresses all of the above.

But note that ddraid is still pre-alpha, because a few bits aren't fully 
hooked up yet, and some interfaces are still in flux.  I'll update the 
tarball to  (ddraid.0.0.6) later tonight, and post the lca slides.

Regards,

Daniel