On Sat, Jun 28, 2003 at 11:54:43PM +0200, Daniel Phillips wrote:
> Yes, now we're on the same page, so to speak. Personally, I don't have much
> interest in working on improving the allocator per se. I'd love to see
> somebody else take a run at that, and I will occupy myself with the gritty
> details of how to move pages without making the system crater.
This sounds like it's behind dependent on physically scanning slabs,
since one must choose slab pages for replacement on the basis of their
potential to restore contiguity, not merely "dump whatever's replaceable
and check how much got freed".
On Sunday 29 June 2003 23:26, Mel Gorman wrote:
>> The point which I am getting across, quite
>> badly, is that by having order0 pages in slab, you are guarenteed to be
>> able to quickly find a cluster of pages to move which will free up a
>> contiguous block of 2^MAX_ORDER pages, or at least 2^MAX_GFP_ORDER with
>> the current slab implementation.
On Sat, Jun 28, 2003 at 11:54:43PM +0200, Daniel Phillips wrote:
> I don't see that it's guaranteed, but I do see that organizing pages in
> slab-like chunks is a good thing to do - a close reading of my original
> response to you shows I was thinking about that.
> I also don't see that the slab cache in its current incarnation is the right
> tool for the job. It handles things that we just don't need to handle, such
> as objects of arbitary size and alignment. I'm sure you could make it work,
> but why not just tweak alloc_pages to know about chunks instead?
A larger question in my mind is what slabs have to do with this at all.
Slabs are for fixed-size allocations. These techniques are explicitly
oriented toward making user allocations more variable, not less.
-- wli
-
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/