Re: sched: Question about big and little cores system with SMP and EAS

[Morten Rasmussen]

On Thu, Jun 17, 2021 at 01:00:12PM +0800, Ley Foon Tan wrote:
> On Wed, Jun 16, 2021 at 9:20 PM Dietmar Eggemann
> <dietmar.eggemann@xxxxxxx> wrote:
> >
> > - Quentin Perret <quentin.perret@xxxxxxx>
> > + Quentin Perret <qperret@xxxxxxxxxx>
> >
> > On 16/06/2021 13:39, Peter Zijlstra wrote:
> > > On Wed, Jun 16, 2021 at 07:29:26PM +0800, Ley Foon Tan wrote:
> > >> Hi all
> > >>
> > >> Would like to ask the experts here regarding the Symmetric
> > >> Multi-Processing mode (SMP) with Energy aware scheduler (EAS) support
> > >> on the big + little cores system.
> > >
> > > And the you ask a question unrelated to either Symmetric MP or EAS :-)
> > >
> > >> Hardware system:
> > >> Big and little cores have almost the same ISA, but the big core has
> > >> some extension instructions that little core doesn't have.
> > >
> > > That problem is unrelated to big.Little / EAS, also by definition that
> > > is not SMP seeing how the 'S' is a blatant lie.
> > >
> > > The simplest solution is to simply disallow usage of the extended ISA
> > > and force mandate the common subset. The complicated answer is something
> > > along the lines of:
> > >
> > >   https://lkml.kernel.org/r/20210608180313.11502-1-will@xxxxxxxxxx
> >
> > We don't encourage asymmetric ISA extensions for EAS*/CAS** on
> > big.Little systems.
> > It would be simply a nightmare to schedule tasks on such systems.
> >
> > The exception to this is the 'asymmetric 32-bit Soc' to support 32bit
> > legacy Apps. The nightmare for scheduling is reduced in this case to CPU
> > affinity, something the task scheduler has to live with already today.
> > (+ DL admission control for 32bit tasks).
> >
> > *  Documentation/scheduler/sched-energy.rst
> > ** Documentation/scheduler/sched-capacity.rst
> 
> Thanks for the reply.
> Yes, forsee it is very complicated and nightmare for software to
> support for SMP mode but HW is not real "symmetric".
> That's why post the question here to ask the advice and comment from
> experts here. So that can feedback to HW team.
> Asymmetric extension instructions issue should more complicated than
> asymmetric 32-bit app support, it can happen in all the areas (kernel,
> app, library and etc).

Indeed. Detecting what extensions a task might use difficult, if not
impossible. Also, we certainly don't want to end up in situation where
the CPU subsets supporting two extensions are disjoint and a task
requires both extensions.

Morten