Scaling of ctdb & DNS round robin and scaling of pdc

[Martin Schwenke]

On Thu, 25 Oct 2012 10:28:55 +0200, Johannes Amorosa | Celluloid VFX
<johannesa at celluloid-vfx.com> wrote:

I haven't seen anyone answer this question directly so I'll make a best
guess and people can correct me...  :-)

> 1. In the ctdb wiki http://wiki.samba.org/index.php/CTDB_Setup it is 
> recommended to use dns round robin for loadbalancing.
> I'm quite new to this - how does samba handle a "session". We need 
> massive I/O for our setup and considering ctdb to avoid our
> smb bottleneck. When is the client getting a "different" IP- Address? 
> Everytime a file is read or stored - our every morning when
> you log into your workstation. If this is the case then loadbalancing 
> would be quite random.

The closest answer is that a client will get a different IP address
when logging into a workstation.  I say "closest" because it doesn't
occur on a per-file basis but I have seen some situations where clients
have connections to multiple nodes.  Samba experts might be able to
explain this more clearly.

Yes, round-robin DNS based load balancing is quite a weak form of load
balancing because it doesn't take into account the actual load on the
nodes.  However, if you have a lot of clients then it can average
out so it looks like some of the clever forms of load balancing.  :-)

> The scaling figures in one of the PDFs are impressive - are these
> "artifical" benchmark tests or real life scaling?

You would probably have to tell us which PDF!  Some of the documents
are pretty old so we would have to find someone who remembers, but
performance has improved over the years rather than degrading.  CTDB is
used to effectively cluster Samba in many places...

peace & happiness,
martin