Christoph Lameter a écrit :
On Thu, 23 Oct 2008, Pekka Enberg wrote:
The problem looks like its freeing objects on a different processor that
where it was used last. With the pointer array it is only necessary
to touch
the objects that contain the arrays.
Interesting. SLAB gets away with this because of per-cpu caches or
because it uses the bufctls instead of a freelist?
Exactly. Slab adds a special management structure to each slab page that
contains the freelist and other stuff. Freeing first occurs to a per cpu
queue that contains an array of pointers. Then later the objects are
moved from the pointer array into the management structure for the slab.
What we could do for SLUB is to generate a linked list of pointer arrays
in the free objects of a slab page. If all objects are allocated then no
pointer array is needed. The first object freed would become the first
pointer array. If that is found to be exhausted then the object
currently being freed is becoming the next pointer array and we put a
link to the old one into the object as well.
This idea is very nice, especially considering that many objects are freed
by RCU, and their rcu_head (which is hot at kfree() time), might be far
away the linked list anchor actually used in SLUB.
At alloc time, I remember I added a prefetchw() call in SLAB in __cache_alloc(),
this could explain some differences between SLUB and SLAB too, since SLAB
gives a hint to processor to warm its cache.
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
