Re: newidle balancing in NUMA domain?

[Nick Piggin <npiggin@suse.de>]

On Mon, Nov 23, 2009 at 06:04:45PM +0100, Ingo Molnar wrote:
> 
> * Nick Piggin <npiggin@suse.de> wrote:
> 
> > > .32 is kind of closed, with us being at -rc8.
> > 
> > It's a bad regression though.
> 
> It's about 3 months too late for that. Ideally we want performance 

Too late for what? Reporting and reverting a regression? I don't
think so. It is not my problem if patches aren't tested well
enough before being merged.

If we release a kernel with this known problematic scheduler behaviour
then it gives userspace application writers far harder targets, and
also it will give *some* 2.6.32 users regressions if we find it has
to be fixed in 2.6.33.


> regressions to be looked for and reported when the patches go into the 
> devel tree. Failing that, -rc1 would be the good time to re-test 
> whatever workload you care about.
> 
> If you cannot test it in a regular fashion you can offload the testing 
> to us, by adding a similar/equivalent workload to 'perf bench sched'. 
> We'll make sure it stays sane.

What exactly tests *were* done on it? Anything that might possibly
trigger its obvious possibility for detremental behaviour? Something
like netperf?