Re: [Xen-devel] RFC: Still TODO for 4.2? xl domain numa memory allocation vs xm/xend

[Stefano Stabellini <stefano.stabellini@xxxxxxxxxxxxx>]

On Fri, 20 Jan 2012, Ian Campbell wrote:
> On Fri, 2012-01-20 at 10:55 +0000, Stefano Stabellini wrote:
> > On Fri, 20 Jan 2012, Ian Campbell wrote:
> > > On Thu, 2012-01-19 at 21:14 +0000, Pasi 
> > > KÃ?Æ?Ã?Æ?Ã?â??Ã?¤rkkÃ?Æ?Ã?Æ?Ã?â??Ã?¤inen wrote:
> > > > On Wed, Jan 04, 2012 at 04:29:22PM +0000, Ian Campbell wrote:
> > > > > 
> > > > > Has anybody got anything else? I'm sure I've missed stuff. Are there 
> > > > > any
> > > > > must haves e.g. in the paging/sharing spaces?
> > > > > 
> > > > 
> > > > Something that I just remembered:
> > > > http://wiki.xen.org/xenwiki/Xen4.1
> > > > 
> > > > "NUMA-aware memory allocation for VMs. xl in Xen 4.1 will allocate
> > > > equal amount of memory from every NUMA node for the VM. xm/xend
> > > > allocates all the memory from the same NUMA node."
> > > 
> > > I'm not that familiar with the NUMA support but my understanding was
> > > that memory was allocated by libxc/the-hypervisor and not by the
> > > toolstack and that the default was to allocate from the same numa nodes
> > > as domains the processor's were pinned to i.e. if you pin the processors
> > > appropriately the Right Thing just happens. Do you believe this is not
> > > the case and/or not working right with xl?
> > 
> > It seems that xend is retrieving numa info about the platform, see
> > pyxc_numainfo, then using those info to pin vcpus to pcpus, see
> > _setCPUAffinity.
> > Still it seems to me more of an hack than the right way to solve the
> > problem.
> 
> Right, so in the absence of any explicit configuration it basically
> picks a NUMA node (via some heuristic) and automatically puts the guest
> into it.
> 
> It seems to me that xl's behaviour isn't wrong as such, it's just
> different.
> 
> I think the important thing is that xl should honour user's explicit
> requests to use a particular node, either via vcpu pinning or cpupools
> etc.
 
Indeed.
Moreover it should be a choice for the allocator and the scheduler to
make, not the toolstack.
Otherwise we end up with non-NUMA algorithms in Xen and NUMA algorithms
in the toolstack?

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