Re: [PATCH 1/3] libxl: check whether vcpu affinity and vnuma info match

[Dario Faggioli <dario.faggioli@citrix.com>]

[-- Attachment #1.1: Type: text/plain, Size: 1708 bytes --]

On Tue, 2015-03-24 at 14:41 +0000, Wei Liu wrote:
> On Tue, Mar 24, 2015 at 02:41:48PM +0100, Dario Faggioli wrote:

> Currently all the checks that returns error are all guest visible
> mis-configurations. I'm trying to reason whether we should return an
> error or just print a warning.
> 
> What is the outcome if you have conflicting setting in vNUMA and vcpu
> affinity? 
>
The outcome is that vcpus will either always run (if hard affinity and
vnuma info mismatch) or prefer to run (if soft affinity and vnuma info
mismatch) on pcpus from pnode(s) different from the pnode specified in
vnuma configuration, and hence where the memory is.

So, for instance, with this:

 vnuma = [ [ "pnode=0","size=1000","vcpus=0-1","vdistances=10,20"  ],
           [ "pnode=1","size=1000","vcpus=2-3","vdistances=20,10"  ] ]

and this:

 cpus_soft = "node:1"
 cpus = "node:0"

in the config file, we have a soft affinity mismatch for vcpus 0,1 and
an hard affinity mismatch for vcpus 2,3.

This means that vcpus 0,1 can run everywhere, but they will prefer to
run on pcpus from pnode 1, while, when the domain was built, they've
been assigned to vnode 0, which has its memory allocated from pnode 0.
This will (potentially) cause a lot of remote memory accesses.

It also means that vcpus 2,3 will only run on pcpus from node:0, while
they've been assigned to vnode 1, which has its memory allocated from
pnode 1. This will mean all memory accesses will be remote memory
accesses

So no functional consequences, but performance will most likely be
affected.

> I guess it's just performance penalty but nothing guest
> visible could happen?
> 
Exactly.

Regards,
Dario

[-- Attachment #1.2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 181 bytes --]

[-- Attachment #2: Type: text/plain, Size: 126 bytes --]

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