Re: [patch 0/2] fix perf. bug in wake-up load balancing for aim7 and db workload

[Ingo Molnar <mingo@elte.hu>]

* Andrew Morton <akpm@osdl.org> wrote:

> "Chen, Kenneth W" <kenneth.w.chen@intel.com> wrote:
> >
> > Here is a respin of the patch with more documentation.
> >
> 
> Thanks.  Can you send me an unwordwrapped version off-list?
> 
> Did I mention "ug"?
> 
> Ingo, what's your plan here?

I really dont like the sysctl hack. Firstly, which precise kernel 
version was tested - do we know that it wasnt e.g. the smpnice 
regression interfering? Secondly, i dont like the sysctl concept itself: 
i really think we should try to find a way for _applications_ to be 
woken up according to their workload.

If we add the sysctl then basically only the benchmarkers will use it - 
99.99% of users will get whatever default we (and distros) provide, and 
the problem wont be solved in any way. In fact, we'll never be able to 
get rid of the knob again i suspect. I'd rather have the wakeup patch 
reverted, and some better method presented. Adding the sysctl just 
removes all the incentive for people to work on solving this problem in 
some real way.

I also refuse to regard this as any sort of emergency that justifies the 
sysctl hack. The test results came clearly late and i suggested to the 
benchmarking guys a long time ago that if they want us to care about 
their workload, and if it's complex to reproduce the benchmark, they 
should distill some simpler test-app for us to so that we can reproduce 
those cases. I'd much rather like to do the simplest thing: revert the 
wakeup patch (we were fine without it for 15 kernel releases), than to 
paper over [permanently!] this particular incarnation of a wider 
problem.

	Ingo