[DRBD-user] Three node cluster?

Arnold Krille arnold at arnoldarts.de
Tue Apr 17 20:03:07 CEST 2012

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.


On Tuesday 17 April 2012 10:08:33 you wrote:
> On Tue, Apr 17, 2012 at 3:05 AM, Arnold Krille <arnold at arnoldarts.de> wrote:
> > On 15.04.2012 22:05, Björn Enroth wrote:
> >> I am looking for information of how to deal with a KVM three node cluster
> >> with DRBD
> >> I have a "baby machine" ubuntu 11.10 pacemaker/drbd cluster with two
> >> nodes,
> >> local disks with drbd setup in between. This is working flawless.
> >> 
> >> My challenge now is that I want to add a third node with the same setup.
> >> How do I handle drbd in this setup? I'd like to have all nodes active, to
> >> be able to migrate resources, mainly kvm virtual guests, around the
> >> cluster
> >> as I see fit. I'd also like pacemaker to be able to dynamically handle
> >> the
> >> load.
> > 
> > While drbd is great, this is exactly our intended use-case and also
> > exactly
> > the reason I am looking at other storage solutions. drbd can't do more
> > than
> > two nodes.
> > 
> > You can of course distribute the drbd-resources so that some are n1/n2,
> > some n2/n3 and some n1/n3, but that becomes an administrators nightmare.
> > And once you decide that you need four nodes with the data present on at
> > least three nodes, you are stuck.
> > You can layer the drbd-resources but that is more meant for semi-distant
> > mirrors and manual fail-over.
> > And if you want live-migrations for your vms with more then two primary
> > filesystem nodes...
> > 
> > I am currently looking at glusterfs, there is also moosefs and ceph(fs),
> > but only the first is meant to be stable enough that redhat gives
> > commercial support for it. There are also other distributed cluster
> > filesystems like lustre, but they lack redundancy.
> 
> FWIW, I agree that GlusterFS is probably the best available option at
> this time, for this use case. I'd recommend Ceph (Qemu-RBD,
> specifically) if the virtualization cluster was larger, but the
> GlusterFS option combines excellent ease-of-use with good redundancy
> and would probably be your best bet for this cluster size.

Afaik ceph can do replication of the kind "give me three replicas, no matter 
how many backend-nodes are up". A feature gluster is missing in the current 
stable version (but I've been told that its coming in 3.3, due next months).

I like the rbd-device interface of ceph, but during my tests with ceph during 
this last winter (yes, the whole week!:) ceph died on me two times and took 
its data with it. glusterfs also has problems when the participating nodes 
change ip-addresses, but still all the data is accessible...

But all these technologies are fairly new, no one yet has found the ultimative 
solution to the split-brain-problem for example, and so your milage may 
vary...

A sidenote to get back on topic: While my co-admin had his small victory some 
months ago, when it was decided to go back from drbd-hosting to his extra-
iscsi machine, last week that machine had a problem starting after an power-
outage. The only vm running fine was that last (and un-important) vm coming 
from drbd...

Have fun,

Arnold
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.linbit.com/pipermail/drbd-user/attachments/20120417/6d0a8c16/attachment.pgp>


More information about the drbd-user mailing list