Date:   Tue, 15 Mar 2022 17:51:41 +0100
From:   Vincent Guittot <vincent.guittot@...aro.org>
To:     王擎 <wangqing@...o.com>
Cc:     Catalin Marinas <Catalin.Marinas@....com>,
        Will Deacon <will@...nel.org>,
        Sudeep Holla <sudeep.holla@....com>,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        "Rafael J. Wysocki" <rafael@...nel.org>,
        Ingo Molnar <mingo@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Juri Lelli <juri.lelli@...hat.com>,
        Dietmar Eggemann <dietmar.eggemann@....com>,
        Steven Rostedt <rostedt@...dmis.org>,
        Ben Segall <bsegall@...gle.com>, Mel Gorman <mgorman@...e.de>,
        Daniel Bristot de Oliveira <bristot@...hat.com>,
        "linux-arm-kernel@...ts.infradead.org" 
        <linux-arm-kernel@...ts.infradead.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH] sched: topology: make cache topology separate from cpu topology

On Tue, 15 Mar 2022 at 02:58, 王擎 <wangqing@...o.com> wrote:
>
>
> >>
> >>
> >> >>
> >> >>
> >> >> >>
> >> >> >>
> >> >> >> >On Thu, 10 Mar 2022 at 13:59, Qing Wang <wangqing@...o.com> wrote:
> >> >> >> >>
> >> >> >> >> From: Wang Qing <wangqing@...o.com>
> >> >> >> >>
> >> >> >> >> Some architectures(e.g. ARM64), caches are implemented below:
> >> >> >> >> cluster:                     ****** cluster 0 *****      ****** cluster 1 *****
> >> >> >> >> core:                         0      1          2      3          4      5           6      7
> >> >> >> (add cache level 1)        c0    c1        c2    c3         c4    c5         c6    c7
> >> >> >> >> cache(Leveln):         **cache0**  **cache1**  **cache2**  **cache3**
> >> >> >> (add cache level 3)        *************share level 3 cache ***************
> >> >> >> >> sd_llc_id(current):     0      0          0      0          4      4           4      4
> >> >> >> >> sd_llc_id(should be): 0      0          2      2          4      4           6      6
> >> >> >> >>
> >> >> >> Here, n always be 2 in ARM64, but others are also possible.
> >> >> >> core[0,1] form a complex(ARMV9),  which share L2 cache, core[2,3] is the same.
> >> >> >>
> >> >> >> >> Caches and cpus have different topology, this causes cpus_share_cache()
> >> >> >> >> return the wrong value, which will affect the CPU load balance.
> >> >> >> >>
> >> >> >> >What does your current scheduler topology  look like?
> >> >> >> >
> >> >> >> >For CPU 0 to 3, do you have the below ?
> >> >> >> >DIE [0     -     3] [4-7]
> >> >> >> >MC  [0] [1] [2] [3]
> >> >> >>
> >> >> >> The current scheduler topology consistent with CPU topology:
> >> >> >> DIE  [0-7]
> >> >> >> MC  [0-3] [4-7]  (SD_SHARE_PKG_RESOURCES)
> >> >> >> Most Android phones have this topology.
> >> >> >> >
> >> >> >> >But you would like something like below for cpu 0-1 instead ?
> >> >> >> >DIE [0     -     3] [4-7]
> >> >> >> >CLS [0 - 1] [2 - 3]
> >> >> >> >MC  [0] [1]
> >> >> >> >
> >> >> >> >with SD_SHARE_PKG_RESOURCES only set to MC level ?
> >> >> >>
> >> >> >> We don't change the current scheduler topology, but the
> >> >> >> cache topology should be separated like below:
> >> >> >
> >> >> >The scheduler topology is not only cpu topology but a mixed of cpu and
> >> >> >cache/memory cache topology
> >> >> >
> >> >> >> [0-7]                          (shared level 3 cache )
> >> >> >> [0-1] [2-3][4-5][6-7]   (shared level 2 cache )
> >> >> >
> >> >> >So you don't  bother the intermediate cluster level which is even simpler.
> >> >> >you have to modify generic arch topology so that cpu_coregroup_mask
> >> >> >returns the correct cpu mask directly.
> >> >> >
> >> >> >You will notice a llc_sibling field that is currently used by acpi but
> >> >> >not DT to return llc cpu mask
> >> >> >
> >> >> cpu_topology[].llc_sibling describe the last level cache of whole system,
> >> >> not in the sched_domain.
> >> >>
> >> >> in the above cache topology, llc_sibling is 0xff([0-7]) , it describes
> >> >
> >> >If llc_sibling was 0xff([0-7] on your system, you would have only one level:
> >> >MC[0-7]
> >>
> >> Sorry, but I don't get it, why llc_sibling was 0xff([0-7] means MC[0-7]?
> >> In our system(Android), llc_sibling is indeed 0xff([0-7]) , because they
> >> shared the llc(L3), but we also have two level:
> >> DIE [0-7]
> >> MC [0-3][4-6]
> >> It makes sense, [0-3] are little cores, [4-7] are bit cores, se only up migrate
> >> when misfit. We won't change it.
> >>
> >> >
> >> >> the L3 cache sibling, but sd_llc_id describes the maximum shared cache
> >> >> in sd, which should be [0-1] instead of [0-3].
> >> >
> >> >sd_llc_id describes the last sched_domain with SD_SHARE_PKG_RESOURCES.
> >> >If you want llc to be [0-3] make sure that the
> >> >sched_domain_topology_level array returns the correct cpumask with
> >> >this flag
> >>
> >> Acturely, we want sd_llc to be [0-1] [2-3], but if the MC domain don't have
> >
> >sd_llc_id refers to a scheduler domain but your patch breaks this so
> >if you want a llc that reflects this topo:  [0-1] [2-3] you must
> >provide a sched_domain level with this topo
>
> Maybe we should add a shared-cache level(SC), like what CLS does:
>
> DIE  [0-7] (shared level 3 cache, SD_SHARE_PKG_RESOURCES)
> MC  [0-3] [4-7]  (not SD_SHARE_PKG_RESOURCES)
> CLS  (if necessary)
> SC    [0-1][2-3][4-5][6-7] (shared level 2 cache, SD_SHARE_PKG_RESOURCES)
> SMT (if necessary)
>
> SC means a couple of CPUs which are placed closely by sharing
> mid-level caches, but not enough to be a cluster.

what you name SC above looks the same as CLS which should not be mixed
with Arm cluster terminology

> >
> >side question, why don't you want llc to be the L3 one ?
>
> Yes, we should set SD_SHARE_PKG_RESOURCES to DIE, but we also want to
> represent the mid-level caches to improve throughput.

so your topology (from a scheduler poV) should be for cpu0:
DIE [0   -       3] [4   -   7] (SD_SHARE_PKG_RESOURCES)
MC  [0 - 1] [2 - 3]             (SD_SHARE_PKG_RESOURCES)
CLS [0] [1]                     (SD_SHARE_PKG_RESOURCES)

so the llc will be the DIE level and load balance will spread tasks in
the different  group of cpus

And regarding EAS, it will look at DIE level for task placement

And of course, this doesn't need any change in scheduler but the
arch_topology.c to return the correct cpumask for your system

>
> Thanks,
> Wang
> >
> >> SD_SHARE_PKG_RESOURCES flag, the sd_llc will be [0][1][2][3]. It's not true.
> >
> >The only entry point for describing the scheduler domain is the
> >sched_domain_topology_level array which provides some cpumask and some
> >associated flags. By default, SD_SHARE_PKG_RESOURCES is set for
> >scheduler MC level which implies that the cpus shared their cache. If
> >this is not the case for your system, you should either remove this
> >flag or update the cpumask to reflect which CPUs really share their
> >caches.
> >
> >>
> >> So we must separate sd_llc from sd topology, or the demand cannot be
> >> met in any case under the existing mechanism.
> >
> >There is a default array with DIE, MC, CLS and SMT levels with
> >SD_SHARE_PKG_RESOURCES set up to MC which is considered to be the LLC
> >but a different array than the default one can be provided with
> >set_sched_topology()
> >
> >Thanks
> >Vincent
> >
> >>
> >> Thanks,
> >> Wang
> >>

Powered by blists - more mailing lists