[PATCH 2/3] arm64: topology: Tell the scheduler about the relative power of cores

[Catalin Marinas]

On Thu, Dec 12, 2013 at 01:36:43PM +0000, Morten Rasmussen wrote:
> On Thu, Dec 12, 2013 at 12:06:49PM +0000, Mark Brown wrote:
> > On Thu, Dec 12, 2013 at 11:35:12AM +0000, Morten Rasmussen wrote:
> > 
> > > I know that this is how it is currently done for ARMv7 and one could
> > > argue that we should do the same for ARMv8 until we have a better
> > > solution. I just want to highlight that setting cpu_power this way is
> > > not generally the right thing to do for big.LITTLE. It will have to be
> > > fixed eventually.
> > 
> > Right, there's a good solid reason for all the work on the scheduler.  I
> > definitely think we ought to be following the same approach on both
> > ARMv7 and ARMv8 to avoid confusion between people based on the platform
> > they're working on.
[...]
> > If you're saying that the current ARMv7 code is always worse than doing
> > nothing then clearly we ought to be removing that code from ARMv7 rather
> > than hurting performance.  I'd been under the impression that what we
> > had there was not ideal but better than nothing in mainline rather than
> > actively harmful.
> 
> For some scenarios it might be better to set cpu_power to reflect the
> relative performance, for others it is worse due to the way cpu_power
> is currently used in the scheduler.
> 
> Setting cpu_power as it is done for v7 may bias the scheduler to put
> heavier tasks on big cpus and will generally put more task on big cpus,
> which is a good thing for some scenarios. However, if you have parallel
> workloads that spawn a worker thread for each cpu and does dynamic work
> distribution in user-space (OpenMP applications for example), then
> setting cpu_power will put two worker threads on some big cpus and leave
> some little cpus idle resulting in slower completion time. It happens on
> TC2.
> 
> We need the code (or something very similar) later when the scheduler
> has been fixed. For v7 it has been left in waiting for that fix, it
> doesn't harm when using the reference big.LITTLE patches. We can do the
> same for v8 to avoid confusion.

If we can't really guarantee the effect of this patch, I would rather
keep it in the LSK kernel only until the scheduler is fixed (can this be
treated as a performance issue independent of the power-aware
scheduling? We could get it merged quicker).

-- 
Catalin