Open Stack

2015-01-09 17:17 GMT+08:00 Sylvain Bauza <sbauza at redhat.com>:

>
> Le 09/01/2015 09:01, Alex Xu a écrit :
>
> Hi, All
>
>  There is bug when running nova with ironic
> https://bugs.launchpad.net/nova/+bug/1402658
>
>  The case is simple: one baremetal node with 1024MB ram, then boot two
> instances with 512MB ram flavor.
> Those two instances will be scheduling to same baremetal node.
>
>  The problem is at scheduler side the IronicHostManager will consume all
> the resources for that node whatever
> how much resource the instance used. But at compute node side, the
> ResourceTracker won't consume resources
> like that, just consume like normal virtual instance. And ResourceTracker
> will update the resource usage once the
> instance resource claimed, then scheduler will know there are some free
> resource on that node, then will try to
> schedule other new instance to that node.
>
>  I take look at that, there is NumInstanceFilter, it will limit how many
> instance can schedule to one host. So can
> we just use this filter to finish the goal? The max instance is configured
> by option 'max_instances_per_host', we
> can make the virt driver to report how many instances it supported. The
> ironic driver can just report max_instances_per_host=1.
> And libvirt driver can report max_instance_per_host=-1, that means no
> limit. And then we can just remove the
> IronicHostManager, then make the scheduler side is more simpler. Does make
> sense? or there are more trap?
>
>  Thanks in advance for any feedback and suggestion.
>
>
>
> Mmm, I think I disagree with your proposal. Let me explain by the best I
> can why :
>
> tl;dr: Any proposal unless claiming at the scheduler level tends to be
> wrong
>
> The ResourceTracker should be only a module for providing stats about
> compute nodes to the Scheduler.
> How the Scheduler is consuming these resources for making a decision
> should only be a Scheduler thing.
>

agreed, but we can't implement this for now, the reason is you described as
below.


>
> Here, the problem is that the decision making is also shared with the
> ResourceTracker because of the claiming system managed by the context
> manager when booting an instance. It means that we have 2 distinct decision
> makers for validating a resource.
>
>
Totally agreed! This is the root cause.


> Let's stop to be realistic for a moment and discuss about what could mean
> a decision for something else than a compute node. Ok, let say a volume.
> Provided that *something* would report the volume statistics to the
> Scheduler, that would be the Scheduler which would manage if a volume
> manager could accept a volume request. There is no sense to validate the
> decision of the Scheduler on the volume manager, just maybe doing some
> error management.
>
> We know that the current model is kinda racy with Ironic because there is
> a 2-stage validation (see [1]). I'm not in favor of complexifying the
> model, but rather put all the claiming logic in the scheduler, which is a
> longer path to win, but a safier one.
>

Yea, I have thought about add same resource consume at compute manager
side, but it's ugly because we implement ironic's resource consuming method
in two places. If we move the claiming in the scheduler the thing will
become easy, we can just provide some extension for different consuming
method (If I understand right the discussion in the IRC). As gantt will be
standalone service, so validating a resource shouldn't spread into
different service. So I agree with you.

But for now, as you said this is long term plan. We can't provide different
resource consuming in compute manager side now, also can't move the
claiming into scheduler now. So the method I proposed is more easy for now,
at least we won't have different resource consuming way between
scheduler(IonricHostManger) and compute(ResourceTracker) for ironic. And
ironic can works fine.

The method I propose have a little problem. When all the node allocated, we
still can saw there are some resource are free if the flavor's resource is
less than baremetal's resource. But it can be done by expose max_instance
to hypervisor api(running instances already exposed), then user will now
why can't allocated more instance. And if we can configure max_instance for
each node, sounds like useful for operator also :)


>
> -Sylvain
>
> [1]  https://bugs.launchpad.net/nova/+bug/1341420
>
>  Thanks
> Alex
>
>
> _______________________________________________
> OpenStack-dev mailing listOpenStack-dev at lists.openstack.orghttp://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
>
>
> _______________________________________________
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-dev/attachments/20150109/fb9da54c/attachment.html>