[Linux-cluster] A few GFS newbie questions: journals, etc

[Lon Hohberger]

On Thu, 2005-06-30 at 11:27 -0400, Andrew Forgue wrote:
> Nate Carlson wrote:
> 
> > On Thu, 30 Jun 2005, Patrick Caulfield wrote:
> >
> >> Not really. The whole point of the quorum is to prevent
> >> "split-brains" where two nodes can see the shared storage but not
> >> each other - which would corrupt the filesystem because they can't
> >> co-ordinate locking.
> >
> 
> This reminds me, how does a cluster behave that has an On-disk quorum
> (Tru64 and Windows, off the top of my head)?  In those OS'es you have to
> dedicate a partition for "quorum".  Are there any advantages to this, or
> am I just misunderstanding something.

Disk based quorum algorithms (with appropriate fencing) can be used to
do N -> 1 failover (e.g. operation with only one node).  They can also
be used to prevent split-brain cases in a more scalable fashion (giving
you 50% votes being okay instead of 51% - this could allow a 4-node
cluster to operate with 2 nodes online, for example).

Of course, now you have to have good shared storage in order for your
cluster to work at all, which isn't a requirement of CMAN or GULM
currently.

-- Lon