Message-ID: <20210412163355.GV6564@kitsune.suse.cz>
Date:   Mon, 12 Apr 2021 18:33:55 +0200
From:   Michal Suchánek <msuchanek@...e.de>
To:     Mel Gorman <mgorman@...hsingularity.net>
Cc:     Vincent Guittot <vincent.guittot@...aro.org>,
        "Gautham R. Shenoy" <ego@...ux.vnet.ibm.com>,
        Michael Neuling <mikey@...ling.org>,
        Vaidyanathan Srinivasan <svaidy@...ux.vnet.ibm.com>,
        Srikar Dronamraju <srikar@...ux.vnet.ibm.com>,
        Rik van Riel <riel@...riel.com>,
        LKML <linux-kernel@...r.kernel.org>,
        Nicholas Piggin <npiggin@...il.com>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Parth Shah <parth@...ux.ibm.com>,
        linuxppc-dev@...ts.ozlabs.org,
        Valentin Schneider <valentin.schneider@....com>
Subject: Re: [RFC/PATCH] powerpc/smp: Add SD_SHARE_PKG_RESOURCES flag to MC
 sched-domain

On Mon, Apr 12, 2021 at 04:24:44PM +0100, Mel Gorman wrote:
> On Mon, Apr 12, 2021 at 02:21:47PM +0200, Vincent Guittot wrote:
> > > > Peter, Valentin, Vincent, Mel, etal
> > > >
> > > > On architectures where we have multiple levels of cache access latencies
> > > > within a DIE, (For example: one within the current LLC or SMT core and the
> > > > other at MC or Hemisphere, and finally across hemispheres), do you have any
> > > > suggestions on how we could handle the same in the core scheduler?
> >
> > I would say that SD_SHARE_PKG_RESOURCES is there for that and doesn't
> > only rely on cache
> >
>
> From topology.c
>
> 	SD_SHARE_PKG_RESOURCES - describes shared caches
>
> I'm guessing here because I am not familiar with power10 but the central
> problem appears to be when to prefer selecting a CPU sharing L2 or L3
> cache and the core assumes the last-level-cache is the only relevant one.

It does not seem to be the case according to original description:

>>>> When the scheduler tries to wakeup a task, it chooses between the
>>>> waker-CPU and the wakee's previous-CPU. Suppose this choice is called
>>>> the "target", then in the target's LLC domain, the scheduler
>>>> 
>>>> a) tries to find an idle core in the LLC. This helps exploit the
This is the same as (b) Should this be SMT^^^ ?
>>>>    SMT folding that the wakee task can benefit from. If an idle
>>>>    core is found, the wakee is woken up on it.
>>>> 
>>>> b) Failing to find an idle core, the scheduler tries to find an idle
>>>>    CPU in the LLC. This helps minimise the wakeup latency for the
>>>>    wakee since it gets to run on the CPU immediately.
>>>> 
>>>> c) Failing this, it will wake it up on target CPU.
>>>> 
>>>> Thus, with P9-sched topology, since the CACHE domain comprises of two
>>>> SMT4 cores, there is a decent chance that we get an idle core, failing
>>>> which there is a relatively higher probability of finding an idle CPU
>>>> among the 8 threads in the domain.
>>>> 
>>>> However, in P10-sched topology, since the SMT domain is the LLC and it
>>>> contains only a single SMT4 core, the probability that we find that
>>>> core to be idle is less. Furthermore, since there are only 4 CPUs to
>>>> search for an idle CPU, there is lower probability that we can get an
>>>> idle CPU to wake up the task on.

>
> For this patch, I wondered if setting SD_SHARE_PKG_RESOURCES would have
> unintended consequences for load balancing because load within a die may
> not be spread between SMT4 domains if SD_SHARE_PKG_RESOURCES was set at
> the MC level.

Not spreading load between SMT4 domains within MC is exactly what setting LLC
at MC level would address, wouldn't it?

As in on P10 we have two relevant levels but the topology as is describes only
one, and moving the LLC level lower gives two levels the scheduler looks at
again. Or am I missing something?

Thanks

Michal

> > >
> > > Minimally I think it would be worth detecting when there are multiple
> > > LLCs per node and detecting that in generic code as a static branch. In
> > > select_idle_cpu, consider taking two passes -- first on the LLC domain
> > > and if no idle CPU is found then taking a second pass if the search depth
> >
> > We have done a lot of changes to reduce and optimize the fast path and
> > I don't think re adding another layer  in the fast path makes sense as
> > you will end up unrolling the for_each_domain behind some
> > static_banches.
> >
>
> Searching the node would only happen if a) there was enough search depth
> left and b) there were no idle CPUs at the LLC level. As no new domain
> is added, it's not clear to me why for_each_domain would change.
>
> But still, your comment reminded me that different architectures have
> different requirements
>
> Power 10 appears to prefer CPU selection sharing L2 cache but desires
> 	spillover to L3 when selecting and idle CPU.
>
> X86 varies, it might want the Power10 approach for some families and prefer
> 	L3 spilling over to a CPU on the same node in others.
>
> S390 cares about something called books and drawers although I've no
> 	what it means as such and whether it has any preferences on
> 	search order.
>
> ARM has similar requirements again according to "scheduler: expose the
> 	topology of clusters and add cluster scheduler" and that one *does*
> 	add another domain.
>
> I had forgotten about the ARM patches but remembered that they were
> interesting because they potentially help the Zen situation but I didn't
> get the chance to review them before they fell off my radar again. About
> all I recall is that I thought the "cluster" terminology was vague.
>
> The only commonality I thought might exist is that architectures may
> like to define what the first domain to search for an idle CPU and a
> second domain. Alternatively, architectures could specify a domain to
> search primarily but also search the next domain in the hierarchy if
> search depth permits. The default would be the existing behaviour --
> search CPUs sharing a last-level-cache.
>
> > SD_SHARE_PKG_RESOURCES should be set to the last level where we can
> > efficiently move task between CPUs at wakeup
> >
>
> The definition of "efficiently" varies. Moving tasks between CPUs sharing
> a cache is most efficient but moving the task to a CPU that at least has
> local memory channels is a reasonable option if there are no idle CPUs
> sharing cache and preferable to stacking.
>
> > > allows within the node with the LLC CPUs masked out. While there would be
> > > a latency hit because cache is not shared, it would still be a CPU local
> > > to memory that is idle. That would potentially be beneficial on Zen*
> > > as well without having to introduce new domains in the topology hierarchy.
> >
> > What is the current sched_domain topology description for zen ?
> >
>
> The cache and NUMA topologies differ slightly between each generation
> of Zen. The common pattern is that a single NUMA node can have multiple
> L3 caches and at one point I thought it might be reasonable to allow
> spillover to select a local idle CPU instead of stacking multiple tasks
> on a CPU sharing cache. I never got as far as thinking how it could be
> done in a way that multiple architectures would be happy with.
>
> --
> Mel Gorman
> SUSE Labs

Powered by blists - more mailing lists