Re: rr_interval experiments

[Nick Piggin <npiggin@suse.de>]

On Fri, Apr 20, 2007 at 10:47:57AM +1000, Con Kolivas wrote:
> On Friday 20 April 2007 01:01, Con Kolivas wrote:
> > This then allows the maximum rr_interval to be as large as 5000
> > milliseconds.
> 
> Just for fun, on a core2duo make allnoconfig make -j8 here are the build time 
> differences (on a 1000HZ config) machine:
> 
> 16ms:
> 53.68user 4.81system 0:34.27elapsed 170%CPU (0avgtext+0avgdata 0maxresident)k
> 
> 1ms:
> 56.73user 4.83system 0:36.03elapsed 170%CPU (0avgtext+0avgdata 0maxresident)k
> 
> 5000ms:
> 52.88user 4.77system 0:32.37elapsed 178%CPU (0avgtext+0avgdata 0maxresident)k
> 
> For the record, 16ms is what SD v0.43 would choose as the default value on 
> this hardware. A load with a much lower natural context switching rate than a 
> kernel compile, as you said Nick, would show even greater discrepancy in 
> these results.
> 
> Fun eh? Note these are not for any comparison with anything else; just to show 
> the effect rr_interval changes have on throughput.

Yeah very interesting, thanks. I was sure that a more modern CPU and/or
one with more cache (in this case, both!) would show bigger differences
even on kbuild.

In this case, 16ms -> infinite results in almost 6% performance
improvement.