Re: Fw: [PATCH] NUMA Slab Allocator

[Manfred Spraul <manfred@colorfullife.com>]

Martin J. Bligh wrote:

>That'd be my inclination .... but OTOH, we do that for pagecache OK.
>
The page cache doesn't have a global hash table.

> Dunno, 
>I'm torn. Depends if there's locality on the file access or not, I guess.
>Is there any *harm* in doing it node local .... perhaps creating a node
>mem pressure imbalance (OTOH, there's loads of stuff that does that anyway ;-))
>
>  
>
The harm is slower kmem_cache_free and a lower hit ratio for the per-cpu 
caches: kmem_cache_free must identify and return wrong node objects, and 
due to these returns, the per-cpu array is more often empty in 
kmem_cache_alloc.

IIRC someone from SGI wrote that they have seen bad performance in 
fork-bomb tests on large cpu count systems which might be caused by 
inter-node traffic on the mm_struct structure and that they think that a 
numa aware allocator would help. As far as I know no tests were done to 
very that assumption.

--
    Manfred