attached is my draft proposal for numa support:

- batching for the return of objects
- the magazine layer as in Bonwicks paper is really missing: it would be 
great to pass filled cpu arrays around between nodes, to reduce the 
amount of list operations.
- global spinlock - I want to get some stats first: the batch return is 
more efficient with a global spinlocks (1 global spinlock, instead of 
(number_of_nodes-1) spinlocks from remote nodes), batch alloc is more 
efficient with a per-node spinlock (one local spinlock instead of one 
global spinlock).

Questions:
- What should happen if a node has no memory? Right now, all memory 
would be considered as foreign.
- What should happen if get_free_pages() returns memory from another 
node? Right now, the code corrupts the object counters.
- is it possible implement
	ptr_to_nodeid()
   on all archs efficiently? It will happen for every kfree().

The patch contains some code duplication, it's possible to merge some 
parts of the cpuarray initialization.

What do you think? The patch boots on uniprocessor, with NUMA simulated 
[replace node==cur_node with !=, etc], but is not yet really tested.

--
	Manfred