[DRBD-user] DRBD and KVM for a HA-Cluster ?

Lentes, Bernd bernd.lentes at helmholtz-muenchen.de
Mon Jan 17 18:56:30 CET 2011

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

JR wrote:

> It has nothing to do with VMs at all.  We're talking about the storage block device (DRBD) being accessible on both nodes simultaneously:
> http://www.drbd.org/users-guide/s-enable-dual-primary.html  I understand "active-active" and "dual-primary" is a common term applicable in many 
> places,but to be clearly, just talking about the configuration options in drbd.conf:

>	  net {
>	    allow-two-primaries;
>	  }
>	  startup {
>	    become-primary-on both;
>	  }


> The reason you make it active-active on both nodes is because with Live-Migration the VM does exist simultaneously on both nodes for a brief period of 
> time as the memory pages/state are
> migrated over and both access the block device simultaneously yet in a synchronized manner through the underlying HyperVisor.  Indeed, there is a
> mechanism in place for synchronization as your >surmised there must be.
> Don't I need a cluster aware FS for that short period of time when both are running ? What mechanism for synchronization do you mean ?
> Can I realise the Live migration with cluster tools, or with KVM tools ? If I use KVM tools, don't they interfere with the HA-Software ?

> They are never "both running" at the same time: though they may exist on both hosts at the same time, one VM instance will stop before the copy of this > VM instance starts from where the original stopped.

> How familiar are you with live migration?  http://www.linux-kvm.org/wiki/images/5/5a/KvmForum2007$Kvm_Live_Migration_Forum_2007.pdf  The key point of
> LiveMigration is that the VM is copied to another physical host while running, after which the original VM is paused, then dirty pages that changed 
> since the initial page copy are synchronized over quickly, and then it does a quick hand off from one physical to host to the other (MAC addresses, etc) 
> before unpausing the recently sync'd VM on the new host.  Check out pages 4-9 in the PDF above for a more detailed description of this process.  While 
> the VM exists on 2 physical hosts for a brief moment during this process, the HyperVisor or KVM makes sure only 1 copy of the VM is active at a time.

> The short answer to your "do I need a cluster aware FS" is: no.  You need block-level synchronization, in your case KVM assures this during a live-
> migration.  These are some fairly semantic low-level details, I just wanted to be precise in what was actually happening here.  DRBD's "allow-two-
> primaries;" is required for LiveMigration.

Live Migration sounds very interesting. I think i will create a dual-primary solution.
Concerning the live migration:
Are they performed with HA-Tools in pacemaker, or with KVM-Tools ? 


More information about the drbd-user mailing list