[Openstack-operators] pci passthrough & numa affinity

Eric Fried openstack at fried.cc
Thu May 24 23:34:06 UTC 2018


How long are you willing to wait?

The work we're doing to use Placement from Nova ought to allow us to
model both of these things nicely from the virt driver, and request them
nicely from the flavor.

By the end of Rocky we will have laid a large percentage of the
groundwork to enable this. This is all part of the road to what we've
been calling "generic device management" (GDM) -- which we hope will
eventually let us remove most/all of the existing PCI passthrough code.

I/we would be interested in hearing more specifics of your requirements
around this, as it will help inform the GDM roadmap.  And of course,
upstream help & contributions would be very welcome.

Thanks,
efried

On 05/24/2018 05:19 PM, Jonathan D. Proulx wrote:
> On Fri, May 25, 2018 at 07:59:16AM +1000, Blair Bethwaite wrote:
> :Hi Jon,
> :
> :Following up to the question you asked during the HPC on OpenStack
> :panel at the summit yesterday...
> :
> :You might have already seen Daniel Berrange's blog on this topic:
> :https://www.berrange.com/posts/2017/02/16/setting-up-a-nested-kvm-guest-for-developing-testing-pci-device-assignment-with-numa/
> :? He essentially describes how you can get around the issue of the
> :naive flat pci bus topology in the guest - exposing numa affinity of
> :the PCIe root ports requires newish qemu and libvirt.
> 
> Thanks for the pointer not sure if I've seen that one, I've seen a few
> ways to map manually.  I would have been quite surprised if nova did
> this so I am poking at libvirt.xml outside nova for now
> 
> :However, best I can tell there is no way to do this with Nova today.
> :Are you interested in working together on a spec for this?
> 
> I'm not yet convinced it's worth the bother, that's the crux of the
> question I'm investigating.  Is this worth the effort?  There's a meta
> question "do I have time to find out" :)
> 
> :The other related feature of interest here (newer though - no libvirt
> :support yet I think) is gpu cliques
> :(https://github.com/qemu/qemu/commit/dfbee78db8fdf7bc8c151c3d29504bb47438480b),
> :would be really nice to have a way to set these up through Nova once
> :libvirt supports it.
> 
> Thanks,
> -Jon
> 
> 
> _______________________________________________
> OpenStack-operators mailing list
> OpenStack-operators at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-operators
> 



More information about the OpenStack-operators mailing list