Re: [Xen-devel] [PATCH 2/3] libxl: automatically set soft affinity after vnuma info

[Dario Faggioli <dario.faggioli@xxxxxxxxxx>]

On Tue, 2015-03-24 at 14:32 +0000, Wei Liu wrote:
> On Tue, Mar 24, 2015 at 02:42:15PM +0100, Dario Faggioli wrote:
> > More specifically, vcpus are assigned to a vnode, which in
> > turn is associated with a pnode. If a vcpu does not have any
> > soft affinity, automatically build up one, matching the pcpus
> > of the said pnode.
> > 
> 
> What about hard affinity? Do we need to generate hard affinity map as
> well? What's your thought on this?
> 
I think we should not do that. Hard affinity has a bigger impact on the
system performance and overall configuration than soft, which is really
just a preference, so I would let the user deal with that explicitly.

For example, setting the hard affinity to the same set of pcpus for
multiple domains could, depending on the system load, prevent some of
(the vcpus of) these domain to be able to execute for quite some time,
while there are idle pcpus in the system. This may well be something the
user wants, but I'd let him say so, if that is the case, rather than
automatically setup things like that for everyone! By only affecting
soft affinity, that doesn't happen, and we stay work conserving by
default, which is important.

Also, only touching soft affinity is exactly what automatic NUMA
placement already does, which in turn is consistent with using the
result of placement to influence the domain's node affinity (i.e., what
Xen 4.3 and 4.4 do, since soft affinity was not there), and it is the
best solution, IMO.

Perhaps we can have another config option, telling whether one wants
vnuma (and/or automatic placement) to affect soft or hard affinity (or
both). I'm not sure how much this would be useful, while it would
certainly make things more complex from both the implementation and the
interface point of view (the latter being the more concerning aspect).
In any case, we can add this at any point in future, if we feel it could
be useful, and I still think the default setting --if such option is not
specified-- should remain "affect soft affinity only".

> The code itself looks good to me.
> 
Thanks for looking at it. :-)

Regards,
Dario

Attachment: signature.asc
Description: This is a digitally signed message part

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel