Attached is my numa aware slab allocator, rediffed against 2.5.46.
It makes the objects that are returned from kmem_cache_alloc strictly 
node local. Unfortunately this means that kmem_cache_free must return 
objects to the home node, which is expensive. (The return is batched, 
but it's still expensive)

I'm not sure that the patch will improve the performance - benchmarks 
are now needed.

TODO:
- implement ptr_to_nodeid() for all archs.The current implementation is 
a dummy, to test the code on non-NUMA systems.
- switch from MAX_NUMNODES to numnodes - Anton proposed that.
- improve the handling of nodes without cpus or without memory.
- add a kmem_cache_alloc_fromnode() function
- replace the kmem_list3 array with an array of pointers, and allocate 
the storage from the right node.
- allocate the head arrays from the node that is local to the cpu that 
accesses the head array.
- check for regressions - I was careful not to undo any cleanups that 
happened between 2.5.42 and 46, but it's possible that I missed some.

--
    Manfred