Re: [PATCH v6 17/21] arch_topology: Limit span of cpu_clustergroup_mask()

[Darren Hart <darren@os.amperecomputing.com>]

On Fri, Jul 08, 2022 at 10:05:32AM +0100, Ionela Voinescu wrote:
> Hi Darren,
> 
> On Thursday 07 Jul 2022 at 17:10:19 (-0700), Darren Hart wrote:
> > On Mon, Jul 04, 2022 at 11:16:01AM +0100, Sudeep Holla wrote:
> > > From: Ionela Voinescu <ionela.voinescu@arm.com>
> > 
> > Hi Sudeep and Ionela,
> > 
> > > 
> > > Currently the cluster identifier is not set on DT based platforms.
> > > The reset or default value is -1 for all the CPUs. Once we assign the
> > > cluster identifier values correctly, the cluster_sibling mask will be
> > > populated and returned by cpu_clustergroup_mask() to contribute in the
> > > creation of the CLS scheduling domain level, if SCHED_CLUSTER is
> > > enabled.
> > > 
> > > To avoid topologies that will result in questionable or incorrect
> > > scheduling domains, impose restrictions regarding the span of clusters,
> > 
> > Can you provide a specific example of a valid topology that results in
> > the wrong thing currently?
> > 
> 
> When CONFIG_SCHED_CLUSTER=y, all typical big.LITTLE platforms will end up
> having a CLS level instead of MC, with an extra flag for the CLS level:
> SD_PREFER_SIBLING. Additional to this, potentially broken cluster
> descriptions in DT (let's say clusters spanning more CPUs than the LLC
> domain) will result in broken scheduler topologies.

You addressed my primary concern below, thank you. Re this point, I was
concerned that we were prioritizing correcting "broken cluster
descriptions" over "correct, but unusual cluster descriptions". Your
solutions seems to elegantly address both.

> 
> This drew our attention that the span of clusters should be restricted
> to ensure they always span less CPUs than LLC, if LLC information exists
> and LLC spans more than 1 core. But the Ampere Altra functionality you
> introduced is maintained. I'll detail this below.
> 
> > > as presented to scheduling domains building code: cluster_sibling should
> > > not span more or the same CPUs as cpu_coregroup_mask().
> > > 
> > > This is needed in order to obtain a strict separation between the MC and
> > > CLS levels, and maintain the same domains for existing platforms in
> > > the presence of CONFIG_SCHED_CLUSTER, where the new cluster information
> > > is redundant and irrelevant for the scheduler.
> > 
> > Unfortunately, I believe this changes the behavior for the existing
> > Ampere Altra systems, resulting in degraded performance particularly
> > latency sensitive workloads by effectively reverting:
> > 
> >   db1e59483d topology: make core_mask include at least cluster_siblings
> > 
> > and ensuring the clustergroup_mask will return with just one CPU for the
> > condition the above commit addresses.
> > 
> 
> It does not change the functionality on Ampere Altra. cpu_coregroup_mask
> will still return 2 CPUs (cluster span). The difference is that
> cpu_clustergroup_mask will see that cpu_coregroup_masks spans the same
> CPUs and it will return a single CPU. This results in the CLS level
> being invalidated, and the MC level maintained. But MC will span 2 CPUs,
> instead of 1, which was the case before your fix. This is alright as
> MC and CLS have the same flags so the existing functionality is fully
> maintained.

Ah, of course. I missed the combined impact of my earlier change plus
yours, which is to first expand MC and then to collapse CLS. It's a
little round about for the Altra, but that seems reasonable as it's a
bit of a corner case in terms topologies.

Thank you for the explanation.

-- 
Darren Hart
Ampere Computing / OS and Kernel