* Ingo Molnar <mingo@elte.hu> wrote:

quoted text
> > The problem is that for each cache, you have an "per-node alien 
> > queues" for each node (see struct kmem_cache nodelists -> struct 
> > kmem_list3 alien). Moving slab metadata to struct page solves this 
> > but now you can only have one "queue" thats part of the same struct.
> 
> yes, it's what i referred to as "distributed, per node cache". It has 
> no "quadratic overhead". It has SLAB memory spread out amongst nodes. 
> I.e. 1 million pages are distributed amongst 1k nodes with 1000 pages 
> per node with each node having 1 page.
> 
> But that memory is not lost and it's disingenous to call it 'overhead' 
> and it very much comes handy for performance _IF_ there's global 
> workload that involves cross-node allocations. It's simply a cache 
> that is mis-sized and mis-constructed on large node count systems but 
> i bet it makes quite a performance difference on low-node-count 
> systems.
> 
> On high node-count systems it might make sense to reduce the amount of 
> cross-node caching and to _structure_ the distributed NUMA SLAB cache 
> in an intelligent way (perhaps along cpuset boundaries) [...]

i think much of this could be achieved by delaying the creation of alien 
caches up until the point a { node X } -> { node Y } alloc/free 
relationship gets established. So if a system is partitioned along 
cpusets, vast areas of the NxN matrix never gets populated with alien 
caches.

	Ingo
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/