nptl perf bench and profiling with pidns patchsets

[Cedric Le Goater]

Cedric Le Goater wrote:
> Kirill Korotaev wrote:
>>> the results were also very reproducible but the profiling was too noisy.
>>> we also changed the kernel. the previous pidns patchset was on a 2.6.21-mm2 
>>> and we ported it on a 2.6.22-rc1-mm1.
>> If reproducible, then were they the same as Pavel posted?

hmm, i don't think I have answered that question clearly. 

No, i didn't get the same results (with a working TSC), even with 
unixbench, and that's why I switched to profiling with ntpl perf. 
Because the difference between these patchsets is so little, I'm 
hoping that there might be one them which could be improved to make 
a real difference. 

Right now, i'm getting better results with suka's by a magnitude of 
1 or 2%, with unixbench and with ntpl perf. but that does not mean 
anything because standard deviation is high and there might be 
scenarii where pavel's patchset is behaving better much better.

So i'll continue studying these pathsets, and run some more tests 
under an unshared pid namespace. If you have any suggestion for 
improvements, please propose. Pavel, could I try your multilevel 
patchset ? 


pavel's proposal is very similar to what we've started talking 
about in 2005 and it fits our requirements. I'm really in favor 
of finishing this pid namespace :)  

thanks,

C.