On 05/20/2013 12:06 AM, Glauber Costa wrote:
> Initial notes:
> ==============
> 
> Please pay attention to new patches that are debuting in this series. Patch1
> changes our unused countries for int to long, since Dave noticed that it wasn't
> being enough in some cases. Aside from that, the major change is that we now
> compute and keep deferred work per-node (Patch13). The biggest effect of this,
> is that to avoid storing a new nodemask in the stack, I am passing only the
> node id down to the API. This means that the lru API *does not* take a nodemask
> any longer, which in turn, makes it simpler.
> 
> I deeply considered this matter, and decided this would be the best way to go.
> It is not different from what I have already done for memcgs: Only a single one
> is passed down, and the complexity of scanning them is moved upwards to the
> caller, where all the scanning logic should belong anyway.
> 
> If you want, you can also grab from branch "kmemcg-lru-shrinker" at:
> 
> 	git://git.kernel.org/pub/scm/linux/kernel/git/glommer/memcg.git
> 
> I hope the performance problems are all gone. My testing now shows a smoother
> and steady state for the objects during the lifetime of the workload, and
> postmark numbers are closer to base, although we do deviate a bit.
> 

Mel, Dave, et. al.

I have applied some more fixes for things I have found here and there as
a result of a new round of testing. I won't post the result here until
Thursday or Friday, to avoid patchbombing you guys. In the meantime I
will be merging comments I receive from this version.

My git tree is up to date, so if you want to test it further, please
pick that up.

I am attaching the result of my postmark run. I think the results look
really good now.