Re: [PATCH] - Memoryless nodes

[Jack Steiner <steiner@sgi.com>]

On Sat, Apr 18, 2009 at 11:00:46AM +0200, Andi Kleen wrote:
> Jack Steiner <steiner@sgi.com> writes:
> 
> > Add support for nodes that have cpus but no memory.
> > The current code was failing to add these nodes
> > to the nodes_present_map.
> 
> That was intentional because a node is defined as an area of memory
> and an area of memory without memory didn't seem to make sense.
> I still don't think it does. Instead the CPUs are assigned
> to nearby nodes. Back then the generic code also couldn't deal
> with them at all, so you might well run into more problems.


Nodes can also be viewed as a collection of cpus and memory that are
located "close together" - ie., same PXM in the SRAT. Normally nodes will
have both memory & cpus but in theory either cpus or memory could be
missing.

I have run a number of tests of nodes with cpus only and so far have not
seen any issues. I certainly won't rule out problems but have not seen
them yet.

If there are any corner cases that you are aware of, let me know & I can
do some more targeted testing.

(Note: this is not a bizarre end case that I'm trying to fix. UV actually
has blades with cpus and no memory available to the OS.)

--- jack