|
Lorn Kay wrote:
I have built a cluster that is something like what you have described
(without CODA) and have found that some of what you want can be done
this way.
Two machines can load share with LVS with one acting as the director (I
used NAT) but things quickly become more complicated when the primary
Director fails and heartbeat has to fire up a new director on the backup
server that only delivers packets locally (or write a whole new script
that doesn't bother with a one-node director). It can be done but it is
more complex than I think it should be. A better solution (IMO) would be
to just use two boxes as directors, and then offer services from inside
the cluster from Real Servers (much easier to maintain clone Real
Servers and grow your cluster that way--see http://systemimager.org/).
For file replication between the servers you can use RSYNC if you don't
have to worry abount file/record locking, but I think a better solution
would be to add a Network Attached Storage (NAS) device to the cluster
(sort of "behind" the Real Servers). Then use NFS between the NAS and
the Real Servers and let NFS take care of file/record locking problems.
(I suppose you could use RAID and NFS on a Linux box as a NAS server
instead of a product from EMC or Network Appliance, etc.) And if you are
worried about a single NAS box use some type of snap-shot method to a
stand-by NAS server.
In my opinion this would help to keep things simple and much easier to
maintain so that each box only has to assume one role: Director, Real
Server or NAS.
Just my 2 cents...
--K
First, to thanks the quick answer.
I also have think in an extra linux box as NAS server, but i've a problem
with that.
I will have users, with their homes in that cluster, so, if the NAS
fails, all cluster 'fail', and i want to avoid that.
Filipe Carvalho
|
