[Sheepdog] Q: stripping / robustness / kvm resume+suspend / ..

[MORITA Kazutaka]

On 12/05/2009 11:29 PM, Daniel Schwager wrote:
> Hi Morita (is this your first name ?),

My first name is Kazutaka.
In Japan, the first name follows the family name.
Please call me Kazutaka. :)

>>> 1. is it possible to use sheepdog as storage for resuming/suspending
>>>    KVM machines ? Normaly, kvm invokes a command like dd or cat to
>>>    store/restore the RAM to/from disk ... this may will not work
>>>    with sheepdog in the current version, or ?
>> Not supported yet, but it will be supported in future.
>> To support resuming/suspending, sheepdog need to support online
>> snapshot. It is included in our TODO list.
> 
> So, i currently cannot take a "normal"-KVM snapshot to a normal
> filesystem (e.g. ext3) 
> using the normale KVM-snapshot approach ? Or do you mean, it's not
> possible
> to store these snapshots directly on the sheepdog-servers.
> 
> May it's (later) possible to 
> 1) create a sheepdog volume manually 
> 2) start kvm-tools manually to store the suspend-data using to the
> volume ?
>    Ok ... there's no sheepdog standalone-client to store data to the
> volume.
>    KVM uses dd/cat to store data - but these commands cannot be used to 
>    access your volume...
> 3) resume the vm using kvm-tools manually (some problem - no standalone
> sheepdog client)
> 4) delete the sheepdog-volume manually

Sorry, I was talking about qemu loadvm/savevm based suspend,
and I don't know well about dd/cat based suspend.
If stand-alone sheepdog client exists, kvm-tools may be able to suspend
sheepdog vm with your approach, I guess. Creating such a client is not
difficult if performance is not cared about.

>>> 2. to be sure - the KVM machines could run on a sheepdog  server
> member,
>>>    but can also run on not-sheepdog-server machine connecting
>>>    to the sheepdog cluster via IP and multicast.
>> It may be possible to support vm runs outside sheepdog servers.
>> We currently assume virtual machines run on sheepdog server machines
>> to enable zero-hop data accessing and to check vm is alive.
>> We can think of removing this restriction if it is needed.
> 
> In "big" (e.g. >10 KVM server) environment, the KVM-server boots 
> via PXE+diskless-NFS (we do ..) and has no or only a 
> small local HDD (using e.g. redhat satellite-server). This will
> reduce overhead of keeping the servers in sync.
>  
> Therefore, it make no sense to add big local storage to the systems.
> It would be great, if you can add this issue 
> "kvm-server must not be a sheepdog-storage server" to your TODO-list (-:

Okay, I think your environment is general and adding it to TODO items is
meaningful. I'll add it.

Thanks,

MORITA Kazutaka