Re: SLUB: Increasing partial pages

[Christoph Lameter]

On Wed, 16 Jan 2008, Matthew Wilcox wrote:

> We tested 2.6.24-rc5 + 76be895001f2b0bee42a7685e942d3e08d5dd46c
> 
> For 2.6.24-rc5 before that patch, slub had a performance penalty of
> 6.19%.  With the patch, slub's performance penalty was reduced to 4.38%.
> This is great progress.  Can you think of anything else worth trying?

Ahhh.. Good to hear that the issue on x86_64 gets better. I am still 
waiting for a test with the patchset that I did specifically to address 
your regression: http://lkml.org/lkml/2007/10/27/245 (where I tried to 
come up with an in kernel benchmark that exposes the issue even more)? It 
is much more likely now that this patchset in mm addresses your regression 
since the hackbench performance improvement fix already reduce it 
partially.

2.6.25 will add the performance enhancements for the free patch that I 
hope will address your issues. In addition the fastpath was reworked. See 
the patch mentioned above.

2.6.26 (hopefully) will remove the per cpu array lookups that are 
currently necessary on x86 and other platforms through improvements to 
the way that percpu data is handled. That leads to further fastpath 
improvements.

There are also ways to optimize the partial block handling further by f.e.
switching the percpu slab on free in order to have a higher fastpath rate.

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