[nova][numa]Question regarding numa affinity balancer/weigher per host

Li Liu liliueecg at gmail.com
Thu Feb 21 16:09:08 UTC 2019


Thanks a lot, Sean for the clarification :P

Regards
Li Liu

On Thu, Feb 21, 2019 at 10:41 AM Sean Mooney <smooney at redhat.com> wrote:

> On Thu, 2019-02-21 at 10:21 -0500, Li Liu wrote:
> > HI Nova folks,
> >
> > I am trying to find out how Numa balance/weighs per host is taking care
> of in Nova.
> >
> > I know how weighers work in general, but it's weighing between hosts. I
> am not so clear when it comes to a single host
> > with multiple sockets, how does nova weigh them?
> the short answer is it doesn't.
>
> we then to pack numa node 1 before we move on to numa node 2
> the actual asignment of numa resouce to the vm is done by the resouce
> tracker on the compute node.
> stephen on cc has done some work around avoidign host with pci device when
> they are not requested
> and im not sur if he extened that to numa nodes as well.
>
> the specifc code in the libvirt dirver and hardware.py file is rather
> complicated and hard to extend
> so while this has come up in the past we have not really put alot of
> errort into this topic.
>
> it is not clear that balancing the vm placement is always corect and this
> is not something a enduser should
> be able to influnce. that means adding a flavor extraspec. this makes
> modeling numa in placement more complicated
> which is another reason we have not really spent much time on this lately.
> one example where you dont want to blance blindly is when you have sriov
> devices, gpus or fpgas on the host.
> in this case assuming you have not already seperated the host into a
> dedicated hostaggrate for host with special deivces
> you would want to avoid blancing so that instance that dont request an
> fpga are first tried to be placed on numa nodes
> without an fpga before they are  assingined to a numa node with an fpga.
> >
> > For instance, a host has 4 sockets, A, B, C, D. When scheduling request
> comes in asking for 4 cores on 2 sockets(2
> > cores per socket), scheduler realized that A+B, A+C, and C+D combination
> can all fit the request. In this case, how
> > does nova make the decision on which combination to choose from?
> if i remember the code correctly it will take resoces for the first two
> numa nodes that can fit the vm.
> >
>
>

-- 
Thank you

Regards

Li
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-discuss/attachments/20190221/9f566e35/attachment-0001.html>


More information about the openstack-discuss mailing list