Xen devel - Aug 2010 - [PATCH] numa: select nodes by cpu affinity

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Good idea. I will take this and rework it a bit and check it in.

 Thanks,
 Keir

On 04/08/2010 13:04, "Andrew Jones" <drjones@redhat.com> wrote:
> 


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Changeset 21913 in http://xenbits.xen.org/staging/xen-unstable.hg

 -- Keir

On 04/08/2010 13:38, "Keir Fraser" <keir.fraser@eu.citrix.com>
wrote:
> Good idea. I will take this and rework it a bit and check it in.
> 
>  Thanks,
>  Keir
> 
> On 04/08/2010 13:04, "Andrew Jones" <drjones@redhat.com>
wrote:
> 
>> 
> 
> 
> 
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

On 08/04/2010 04:38 PM, Keir Fraser wrote:
> Changeset 21913 in http://xenbits.xen.org/staging/xen-unstable.hg
> 
>  -- Keir
> 
> On 04/08/2010 13:38, "Keir Fraser"
<keir.fraser@eu.citrix.com> wrote:
> 
>> Good idea. I will take this and rework it a bit and check it in.
>>
>>  Thanks,
>>  Keir
>>
>> On 04/08/2010 13:04, "Andrew Jones"
<drjones@redhat.com> wrote:
>
I also considered managing the nodemask as new domain state, as you do,
as it may come in useful elsewhere, but my principle of least patch
instincts kept me from doing it...

I''m not sure about keeping track of the last_alloc_node and then always
avoiding it (at least when there''s more than 1 node) by checking it
last. I liked the way it worked before, favoring the node of the
currently running processor, but I don''t have any perf numbers to know
what would be better.

I''ve attached a patch with a couple minor tweaks. It removes the
unnecessary node clearing from an empty initialized nodemask, and also
moves a couple of domain_update_node_affinity() calls outside
for_each_vcpu loops.

Andrew


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

On 04/08/2010 17:01, "Andrew Jones" <drjones@redhat.com> wrote:
> I also considered managing the nodemask as new domain state, as you do,
> as it may come in useful elsewhere, but my principle of least patch
> instincts kept me from doing it...
Yeah, I don''t fancy iterating over all vcpus for every little bitty
allocation. So it''s a perf thing mainly for me.
> I''m not sure about keeping track of the last_alloc_node and then
always
> avoiding it (at least when there''s more than 1 node) by checking
it
> last. I liked the way it worked before, favoring the node of the
> currently running processor, but I don''t have any perf numbers to
know
> what would be better.
Well, you can expect vcpus to move around within their affinity masks over
moderate timescales (like say seconds or minutes). And in fact the original
credit scheduler *loves* to migrate vcpus around the place over much less
reasonable timescales than that (sub second). It is nice to balance our
allocations rather than hitting one node ''unfairly'' hard.
> I''ve attached a patch with a couple minor tweaks. It removes the
> unnecessary node clearing from an empty initialized nodemask, and also
> moves a couple of domain_update_node_affinity() calls outside
> for_each_vcpu loops.
Thanks, I tweaked your tweaks (just one tiny optimisation) and applied it so
it should show up in the staging tree rsn.

 -- Keir
> Andrew


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel