[DRBD-user] DRBD, Xen, HVM and live migration
arnold at arnoldarts.de
Fri Sep 23 22:20:33 CEST 2011
On Friday 23 September 2011 21:37:51 you wrote:
> On 09/23/11 19:03, Arnold Krille wrote:
> > On Friday 23 September 2011 18:28:41 Felix Frank wrote:
> >> On 09/23/2011 06:18 PM, Bart Coninckx wrote:
> >>> Felix,
> >>> do you combine your setup with clustering software? Or do you script a
> >>> live migration? I'm intrigued by how you handle the timing in the
> >>> latter case.
> >> ah, I didn't make this really clear.
> >> I haven't set this up with either live migration capabilities nor
> >> pacemaker. All I have is a hot standby for manual recovery in case of a
> >> primary node failure. Sorry.
> >> I believe that for pacemaker, you would have both your nodes be primary
> >> at all times (i.e., not use a master-slave set at all).
> >> Not a thrilling perspective, I guess.
> > Dual primary isn't really needed (but its very nice to have).
> > If you use one drbd-resource per vm, you need the dual-primary only
> > during the migration.
> > You can also run single-master that is exported via iscsi/aoe and mounted
> > to all nodes, cluster filesystem on top and you also get an extendible
> > vm- platform. But then you will have vms writing over network to the
> > other node through iscsi which is writing back to your node via drbd...
> That is an interesting idea, though it brakes a bit the high
> availability mission: if one node goes down, you would need to connect
> to the same node (local iSCSI session). I'm told this is not good. It
> would be an ideal situation though, except for the additional network
> cards (seperation iSCSI and DRBD traffic).
True, local iscsi is not nice. But you can get around this by booting your
machines from an gpxe-iso/floppy and tell them the iscsi-device via dhcp. That
is actually a pretty cool solution, it scales to more then two nodes rather
easily. And when you decide that a vm needs to many resources, you can always
adopt the mac-addresses in your dhcp-config and boot that iscsi-image on a real
machine. The "problem" with this solution is that it has that double-network-
traffic penalty when a vm runs on the drbd-slave, it will send both read- and
write-requests across the network two times and you loose the advantage of
drbds local reading. But when your system get big enough, you will probably
not run the vm's on the drbd machines to optimize performance...
Have a nice weekend,
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 198 bytes
Desc: This is a digitally signed message part.
More information about the drbd-user