[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] PV-vNUMA issue: topology is misinterpreted by the guest



On 07/22/2015 11:49 AM, Dario Faggioli wrote:
On Wed, 2015-07-22 at 11:32 -0400, Boris Ostrovsky wrote:
On 07/22/2015 10:50 AM, Dario Faggioli wrote:
Yep. Exacty. As Boris says, this is a generic scheduling issue, although
it's tru that it's only (as far as I can tell) with vNUMA that it bite
us so hard...
I am not sure that it's only vNUMA. It's just that with vNUMA we can see
a warning (on your system) that something goes wrong. In other cases
(like scheduling, or sizing objects based on discovered cache sizes) we
don't see anything in the log but system/programs are making wrong
decisions.

I'm not questioning that the guest scheduler is put in a position where
it may make weird and/or wrong choices... Let's put it like this: on
that very test box, or any other NUMA box I've worked on, I've never
seen performance affected _so_much_ as in the vNUMA case.

In fact, of course there are other issues (like the ones you're
mentioning, caused by this), but it's only with vNUMA that I see 2 out
of 4 vcpus completely lost! :-/

My guess would be that scheduling domains are not set up properly. And perhaps because your test has lots of IO the load doesn't cross thresholds that are needed to switch to another domain?

Does it work the same way if you run a compiled busy loop (i.e. 'main(){while(1);}')?


(And your results above may well be the example of that)

Right. BTW, the example was meant at investigating what you suggested,
i.e., to have the guest topology follow the host topology. I tried, but
could not achieve it. Do you think I'm doing something wrong? May it be
that it's not always doable (ISTR Andrew saying that it is also the pcpu
where the vcpus are created that matters)?

I don't think this is currently doable with what we have for CPUID support in xl syntax. I am pretty sure we need to at least be able to specify all leaf 4's indexes. And we can't.

BTW, irrespective of this particular problem, adding support for indexed CPUID leaves would be a good idea.

-boris


I'm (re)asking because, if you think that's feasible, I can work on
that. I fit's not, well... :-/


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


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.