Open Stack

> From: Thomas Thrainer <thomasth at google.com>
> To: openstack at lists.openstack.org, 
> Date: 01/09/2014 04:51 AM
> Subject: [Openstack] Force volume on same host as VM
> 
> Hi,
> 
> I try to figure out how to force cinder volumes on the same host as 
> the VM it is attached to (or will be attached to). OTOH, I would 
> still like to take advantage of nova-scheduler for automatic host 
> selection for VMs. I don't want to use an availability zone per 
> host, because then the host selection would have to be done by the 
> API client. So I'm wondering what the best solution for this use-
> case might be.
> 
> While researching, I stumbled upon topology aware placement [0] 
> which seems interesting, but I couldn't find any further information
> about the status of it. Also, there seems to be some effort going 
> into building a central scheduler instead of having nova-scheduler 
> and cinder-scheduler [1][2], but again, I couldn't figure out what 
> the current status of this is.
> 
> So, my question is if this use-case will be better supported in the 
> future (Icehouse maybe already?), and how a workaround could look 
> like for now.
> 
> Thanks a lot,
> Thomas
> 
> [0]: https://etherpad.openstack.org/p/HavanaTopologyAwarePlacement
> [1]: https://www.mail-archive.com/openstack-dev@lists.openstack.org/
> msg07981.html
> [2]: http://openstack.10931.n7.nabble.com/cinder-metrics-with-nova-
> scheduler-tt22974.html#a23038

As long as you schedule one thing after another, you might miss the best 
answer.  Some of us are advocating getting to the point where one 
scheduler makes a simultaneous decision about a whole group of resources 
of various kinds.  There was some discussion of the general idea at the 
Icehouse summit

https://etherpad.openstack.org/p/NovaIcehouse-Smart-Resource-Placement

The general issue is widely recognized and at least somewhat appreciated, 
and not something we will be able to do in one cycle.  We suggested a 
roadmap that starts with making Nova able to make a decision about a 
heterogenous group of VMs all at once

https://etherpad.openstack.org/p/NovaIcehouse-Instance-Group-API

and a relatively unambitious version (no nested grouping, relatively 
familiar API phasing) of that was the recommended outcome.

Two big changes were planned to happen to Nova first.  One is 
incorporating Boris' suggestion to make the schedulers keep a copy of the 
relevant hypervisor state in local memory.  The other is a structural 
change to separate out the Nova scheduler (without increasing its 
ambitions); this is now known as Gantt.  Those two should complete before 
the instance group work is done.  Given where we are on the calendar right 
now, I am not so sure how much instance group work will get done in 
Icehouse.

Regards,
Mike
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack/attachments/20140115/a85ff4f6/attachment.html>