[one-users] Pre-migrate and post-migrate actions

[Székelyi Szabolcs]

Hi,

any news on this? Actually this issue is blocking our progress in a project, 
so we'd really like to hear your opinion. Maybe I forgot to write that we'd 
willingly fix this ourselves and then contribute the changes back to 
OpenNebula. We're seeking your advice to make sure that the work we put into 
this conforms to your plans, standards and requirements.

Note that we ran into the same problem recently in relation to setting up the 
ebtables-based network for VMs on live migration. [0]

Thanks,
-- 
cc

[0] http://lists.opennebula.org/pipermail/users-opennebula.org/2010-
July/002303.html

On Wednesday 23 June 2010 20.09.23 Székelyi Szabolcs wrote:
> Hello,
> 
> we faced a problem and thought about more than one ways to fix it,
> unfortunately they require source modification. We would like to know your
> opinion.
> 
> The problem is that we'd like to use iSCSI targets to store the VMs' disks,
> one target per VM. The idea is that if a VM starts running on a host, then
> the host logs in to the corresponding iSCSI target, thus sees the disk
> images for that VM.
> 
> This works now, we're able to do this with a custom TM driver. The problem
> is with live migration. Unfortunately we found no possibility to make ONE
> call a trigger that logs in to the target on the destination host right
> before live migration starts, which means the VM is unable to start on the
> new host.
> 
> The most clear solution seems to be to make LCM call not just the VMM to do
> the migration, but the TM driver beforehand [1] [2]. Due to the
> asynchronous nature of LCM, new VM states must be introduced, just like as
> it is done now with non-live migration.
> 
> Another possibility is to hack VirtualMachineManager or
> VirtualMachineManagerDriver to somehow run the appropriate commands before
> (and after) migration, before returning the result to the LCM.
> 
> The third and most ugly way would be to hack one_vmm_kvm.rb (yes, we're
> using KVM) to do the job.
> 
> What is your opinion about the way to implement this?
> 
> Thanks,