[Linux-cluster] newbie: gfs merge

[Bowie Bailey]

wolfgang pauli wrote:
> > > 
> > > Hm... Thanks for you answer! I am definetelly confused a bit. Even
> > > after reading you post of last week. I understand that i can not
> > > merge the file systems. Our setup is very basic. We have to linux
> > > machines who could act as file server and we thought that we could
> > > one (A) have working as an active backup of the other (B). Is that
> > > what the documentation calls a failover domain, with (B) being the
> > > failover "domain" for (A)? Until now, we were running rsync at
> > > night, so that if the first of the two servers failed, clients
> > > could mount the NFS from the other server. There is nothing fancy
> > > here, like a SAN I guess, just machines connected via ethernet
> > > switches. So basically the question is, whether it is possible to
> > > keep the filesystems on the two servers in total sync, so that it
> > > would not matter whether clients mount the remote share from (A)
> > > or (B). Whether the clients would automatically be able to mount
> > > the GFS from (B), if (A) fails.
> > 
> > No, GFS doesn't work quite like that.  What you have is something
> > more like this:  Two machines, (A) and (B), are file servers.  A
> > third machine, (C), is either a linux box exporting it's filesystem
> > via GNBD, or a dedicated storage box running iSCSI, AoE, or
> > something similar that will allow multiple connections.  (A) and
> > (B) are both connected to the GFS filesystem exported by (C).  If
> > either (A) or (B) goes down, the other one can continue serving the
> > data from (C).  They don't need to be synchronized because they are
> > using the same physical storage.  And, if the application permits,
> > you can even run them both simultaneously. 
> > 
> > You are looking for something different.  There is a project out
> > there for that, but it is not production ready at this point. 
> > Maybe someone else remembers the name. 
> 
> Oh, OK. This would makes sense to me. But I still have some
> questions.. 
> 
> 1. Would this reduce the load on (C)?

Reduce it from what?  (C) would be a completely different type of
machine from (A) and (B).  (A) and (B) are application systems, while
(C) is just a fileserver.  (C) would not need to be quite as fast as
the others, just fast enough to keep up with the I/O requirements of
the storage and the GFS/Cluster overhead.

> 2. I know how to export the gfs from (C) and mount it on (A) and (B),
> but how to the clients know whether they should connect to (A) or
> (B). Is this managed my clvmd?

No, this is managed by your network.  If (A) and (B) are running the
same software, it doesn't matter which one they connect to.  On my
system, I have a Foundry ServerIron that load-balances the two
machines.  You can also do it using LVS software, such as the stuff in
the Linux HA project.

-- 
Bowie