[Openstack-devel] Volatile images and/or hybrid cloud use cases

[Thomas Goirand]

On 02/08/2013 04:19 PM, Daniel Pocock wrote:
> On 07/02/13 17:58, Thomas Goirand wrote:
>> On 02/08/2013 12:55 AM, Thomas Goirand wrote:
>>   
>>> On 02/07/2013 02:36 AM, Daniel Pocock wrote:
>>>     
>>>> I'm just wondering if anyone can comment on the following two topics:
>>>>
>>>>
>>>> a) volatile images - e.g. development servers where new packages are
>>>> installed every few days and where the entire system state must be
>>>> preserved on every reboot.  Can the whole image concept be bypassed and
>>>> the entire instance runs from persistent storage?
>>>>       
>>> The feature that you are looking for is called "boot from volume". Eg,
>>> Cinder provides a block storage to your nova VM, which boots from it.
>>> Then it's more or less like a VPS.
>>>
>>>     
> 
> This first option may eliminate the need to worry about the second
> option (below), as it suggests that all the existing VPSes could simply
> be converted to volumes under the control of Cinder.  Does that make sense?

Yes it does.

> When I went looking for details, I found one bug, which looks like it is
> just an inconvenience, booting such a VM requires a (real or dummy)
> image to exist in Glance even if it is not used:
> https://bugs.launchpad.net/nova/+bug/1008622

This bug is from 2012-06-04, and it looks like it has been fixed. So if
it's still present in Essex/Wheezy (which I'm not sure about), then it
must have been fixed in Folsom.

>>> To run Openstack with Xen, you need XAPI, otherwise called "XCP", which
>>> I happen to also maintain in Debian (it's been in Wheezy for more than a
>>> year now). So, if you are currently running Xen using the python-based
>>> daemon with the "xm" commands, then you will need to switch to something
> 
> This is the use case I'm thinking of - people who have a collection of
> VMs created in the days of lenny or squeeze and want to start
> introducing OpenStack to their environment.

Forget it, it wont work! Unless someone writes the necessary libvirt
stuff so that we can use libvirt+xen. But currently, I haven't even
packaged this use case (nova-compute-xen is for XAPI, not for libvirt+xen).

> Everything I read about OpenStack seems to be aimed at greenfield
> projects with a budget to buy new servers.

Well, what's the point of running cloud computing if you only run few
virtual machines?

>>> As for the state of XCP in Debian, it seems that people tend to run more
>>> CentOS than Debian + XCP, and Citrix is pushing hard for their CentOS
>>> appliance CD image rather than the Debian flavor. It's been more than 6
>>> months they released XCP 1.6 (as a CentOS based ISO image), but nothing
>>> is coming to update what we have in Debian. Probably this will change
>>> later on if Citrix pushes to unify the source code they have: the Debian
>>> package carries 42 patches !!! Most of these are of course written
>>> directly from upstream, but that shows how much they need to rework
>>> their CentOS based code so that it understand better how Debian works.
>>> One of the thing I worked hard on last year was fixing all these
>>> redhat-ismes that were embedded in upstream source code.
>>>
>>>     
> The installations I am thinking about right now are all Xen, so this is
> something I need to look at in more depth.
> 
> Could you comment on the same problem from a KVM perspective?  E.g. if
> somebody has a vanilla KVM environment (or they convert their Xen
> environment to KVM), then they can bring in OpenStack and start running
> Nova on some of those boxes?

I don't see why it would be a problem. Though it seems to be it would be
rather messy to do that. Why wouldn't you just convert your currently
existing VMs into volumes for Cinder, so that you can manage them on the
cloud?

>> Oh, and I forgot. Yes, it seems you can actually run Openstack and still
>> have XAPI based virtual machines... It's in fact the way to go to setup
>> Openstack with XCP: you would run a virtual machine that holds the
>> openstack compute daemon, which talks to XAPI through it's plugin.
>
> Is such a hybrid setup likely to cause other headaches though?
> 
> For example, when OpenStack adjusts iptables or ethernet bridge
> configuration, does that have any side-effects for existing VMs or VPSes
> on the same box?  Or they can happily co-exist and it won't change their
> networking?

Right, I didn't think about networking. As much as I understand,
Openstack would use the xapi bridge interfaces, just like XCP, so
probably it would work. This would need more testing though, and I would
*never* try this on a production server.

> I've had a good look at the proposition from Ubuntu, they make it all
> look very easy on their quick start guide and their marketing materials
> and it would be good to see a Debian alternative to this.

If someone tells you that the Ubuntu packages are easy to use, don't
trust them. :)

There's *a lot* of things that are needed to do by hand, which is quite
error prone.

> I think that
> demonstrating a migration path for existing Debian users will be a
> helpful part of this, because it will bring more Debian-minded people
> into the OpenStack community.

As much as I would like to agree, I don't! :)

I do think that if you plan on using Openstack, then you got to think
about setting it up from scratch, and this, for many reasons. The
biggest concern being networking.

Typically, you'd have one Quantum server that would act as a gateway,
plugged on one interface to the public internet, with the other
interface connecting to other physical server on a private LAN. VMs are
then given public IP addresses using GRE tunnels on the LAN (so that
each tenant can have his private VLAN). This alone may bring lots of
troubles if you plan on doing an hybrid setup. So I wouldn't recommend
this path.

I hope that helps,
Cheers,

Thomas