Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Tue, 14 Sep 2021 13:58:03 +0200
From:   Peter Zijlstra <peterz@infradead.org>
To:     Vincent Guittot <vincent.guittot@linaro.org>
Cc:     Yicong Yang <yangyicong@hisilicon.com>,
        Ingo Molnar <mingo@redhat.com>,
        Juri Lelli <juri.lelli@redhat.com>,
        Dietmar Eggemann <dietmar.eggemann@arm.com>,
        Steven Rostedt <rostedt@goodmis.org>,
        Ben Segall <bsegall@google.com>, Mel Gorman <mgorman@suse.de>,
        Daniel Bristot de Oliveira <bristot@redhat.com>,
        21cnbao@gmail.com,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        "Song Bao Hua (Barry Song)" <song.bao.hua@hisilicon.com>,
        prime.zeng@huawei.com,
        "guodong.xu@linaro.org" <guodong.xu@linaro.org>
Subject: Re: [RFC] Perfomance varies according to sysctl_sched_migration_cost
Message-ID: <YUCOSyFITy/dHcJI@hirez.programming.kicks-ass.net>
References: <ef3b3e55-8be9-595f-6d54-886d13a7e2fd@hisilicon.com>
 <CAKfTPtDJip17rdZ0CMDwUztE7BGUF+J_O2Up=S5GLYrqtkA67A@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <CAKfTPtDJip17rdZ0CMDwUztE7BGUF+J_O2Up=S5GLYrqtkA67A@mail.gmail.com>
Precedence: bulk

On Tue, Sep 14, 2021 at 11:04:03AM +0200, Vincent Guittot wrote:
> 
> I would say that it's a heuristic value that works for most of system
> but it should probably be tuned per platform. But also note that it's
> quite difficult to get a correct value
> 

Right; so back before CFS there was some boot time benchmarks that
measured something for each sched domain.

Conceptually that makes sense, the larger the domain, the larger the
cost, also, you get per platform etc..

In practise it had boot to boot variance and virt fail written all over
it, which is why Ingo ripped it out. I think someone once tried to bring
some of it back, but that was a long time ago.

I'm also not convinced boot time benchmarks are the best idea, because
the above reasons, but perhaps we can do something topology based, and
maybe using a few platform inputs.

And as with anything, some benchmarks will like it, others will not like
it. It's only worth the complexity if we can get an improvement across
the board.