[Linux-cluster] GNBD, how good it is ?

[Benjamin Marzinski]

On Sat, Jul 10, 2004 at 08:38:57PM +0200, Kresimir Kukulj wrote:
> Quoting Benjamin Marzinski (bmarzins at redhat.com):
> > On Fri, Jul 09, 2004 at 03:53:00PM +0100, Gareth Bult wrote:
> > > :)
> > > 
> > > I do appreciate all that, however there are some press releases out
> > > there that are not so clear ..
> > > 
> > > There is certainly an implication in the news items I've seen that this
> > > is "THE GFS" code .. as opposed to being a new and unstable version ..
> > > 
> > > .. Incidentally, I was being kind - I've had many kernel crashes, even
> > > after getting it going ..
> > > 
> > > Regards,
> > > Gareth.
> > 
> > The code you are using is not the code currently being sold by redhat.
> > That is the 6.0 code. You can download that in SRPM form at
> > ftp://ftp.redhat.com/pub/redhat/linux/enterprise/3/en/RHGFS/i386/SRPMS/
> > 
> > There is fairly complete documentation for this code. However it does not use
> > the DLM.  Instead, GULM handles all the cluster manager issues.  This code only
> > runs on 2.4 kernels.
> > 
> > The CVS code is going to be sold starting with RHEL 4. Some of the components,
> > like the dlm are just now gotten out of the development stage. Others, like
> 
> Is this new DLM still dependent on single lock storage or is it distributed
> (like in OpenDLM) ?
> 
> > gnbd have been drastically rewritten.  We REALLY appreciate all the testing
> 
> You are saying that GNBD is rewritten. How does it compare to GNBD in
> GFS-6.0 (version sold by RedHat) in stability, performance, features ?

Since it has just been rewritten, it is currently pretty unstable...
The rewrite involved removing large chunks of gnbd from the kernel, and
doing them in user space.  This should make it easier to maintain. So once
it stabilizes, it should be better.

Performance testing will be done as soon as I'm happy with the stability.
It should be pretty much the same... Not too much of the core functionality
changed.

The features are pretty much identical, except that the new code auto
reconnects if it looses a connection.

The rewrite was done because the block layer had some largish changes from
2.4 to 2.6, and to be more inline with the way redhat ships maintains
products.

-Ben Marzinski
bmarzins at redhat.com