Re: VM balancing issues on 2.6.13: dentry cache not getting shrunk enough

["Martin J. Bligh" <mbligh@mbligh.org>]

>> > Second is Sonny Rao's rbtree dentry reclaim patch which is an attempt
>> > to improve this dcache fragmentation problem.
>> 
>> FYI, in the past I've tried this patch to reduce dcache fragmentation on
>> an Altix (16k pages, 62 dentries to a slab page) under heavy
>> fileserver workloads and it had no measurable effect. It appeared
>> that there was almost always at least one active dentry on each page
>> in the slab.  The story may very well be different on 4k page
>> machines, however.
> 
> I always thought dentry freeing would work much better if it
> was turned upside down.
> 
> Instead of starting from the high level dcache lists it could
> be driven by slab: on memory pressure slab tries to return pages with unused 
> cache objects. In that case it should check if there are only
> a small number of pinned objects on the page set left, and if 
> yes use a new callback to the higher level user (=dcache) and ask them
> to free the object.
> 
> The slab datastructures are not completely suited for this right now,
> but it could be done by using one more of the list_heads in struct page
> for slab backing pages.
> 
> It would probably not be very LRU but a simple hack of having slowly 
> increasing dcache generations. Each dentry use updates the generation.
> First slab memory freeing pass only frees objects with older generations.

If they're freeable, we should easily be able to move them, and therefore 
compact a fragmented slab. That way we can preserve the LRU'ness of it.
Stage 1: free the oldest entries. Stage 2: compact the slab into whole
pages. Stage 3: free whole pages back to teh page allocator.

> Using slowly increasing generations has the advantage of timestamps
> that you can avoid dirtying cache lines in the common case when 
> the generation doesn't change on access (= no additional cache line bouncing)
> and it would easily allow to tune the aging rate under stress by changing the 
> length of the generation.

LRU algorithm may need general tweaking like this anyway ... strict LRU
is expensive to keep.

M.
-
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/