Re: [Xen-devel][PATCH]pcpu tuples [was Re: [Xen-devel] Xen 3.4.1 NUMA support]

[Andre Przywara <andre.przywara@xxxxxxx>]

Dulloor wrote:
> Attached is a patch to construct pcpu tuples of the form
> (node.socket.core.thread), and (currently)used by xm vcpu-pin utility.
Without having looked further at the patch: There will be problems with 
that notation. The assumption that one node consist of at least one 
socket is no longer true with AMD's upcoming Magny-Cours processors, 
which features _two_ nodes in one socket.
The socket information as it is of interest for licensing purposes and 
for the voltage domains. I suppose that power aware scheduling is out of 
scope for the current scheduler, so we could ignore the socket 
information here at all.
Shared cache would be an interesting information to consider for 
scheduling purposes, but again here the socket information is 
misleading, as each node of the Magny-Cours processor has it's own L3 
cache, there is no cache shared across the two nodes in one package.
Xen already detects the NUMA topology on the new system correctly:
nr_cpus                : 48
nr_nodes               : 8
cores_per_socket       : 12
threads_per_core       : 1
I don't know details about the usual IA64 topology, though.

I see currently these possible topologies for x86-64 systems:
Core2-based: 1 (fake) node, n sockets
AMD64/Nehalem: n nodes, 1 socket/node
AMD G34: n nodes, 2 or 1 nodes/socket(!)
That looks like that it will not be easy to combine all of those. One 
possibility would be to join nodes and sockets into one entity (use 
sockets on older systems (L2 cache domains) and nodes on AMD/newer Intel 
systems (memory controller domains)). But I don't have a handy name for 
that beast (left alone nockets ;-)

Although it can be quite useful to have such a notation, I am not sure 
whether it will really help. Eventually you want to go away from manual 
assignment (and be it at domain's runtime via "xm vcpu-pin").
Looking forward to any comments.

Regards,
Andre.

> -dulloor
>
> On Fri, Nov 13, 2009 at 11:02 AM, Keir Fraser <keir.fraser@xxxxxxxxxxxxx> wrote:
> > On 13/11/2009 15:40, "Ian Pratt" <Ian.Pratt@xxxxxxxxxxxxx> wrote:
> >
> > > > Even better would be to have pCPUs addressable and listable explicitly as
> > > > dotted tuples. That can be implemented entirely within the toolstack, and
> > > > could even allow wildcarding of tuple components to efficiently express
> > > > cpumasks.
> > > Yes, I'd certainly like to see the toolstack support dotted tuple notation.
> > >
> > > However, I just don't trust the toolstack to get this right unless xen has
> > > already set it up nicely for it with a sensible enumeration and defined
> > > sockets-per-node, cores-per-socket and threads-per-core parameters. Xen should
> > > provide a clean interface to the toolstack in this respect.
> > Xen provides a topology-interrogation hypercall which should suffice for
> > tools to build up a {node,socket,core,thread}<->cpuid mapping table.
> >
> >  -- Keir
> >
> >
> >
> > _______________________________________________
> > Xen-devel mailing list
> > Xen-devel@xxxxxxxxxxxxxxxxxxx
> > http://lists.xensource.com/xen-devel

--
Andre Przywara
AMD-OSRC (Dresden)
Tel: x29712


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