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, 5 Oct 2021 11:24:23 +0200
From:   Peter Zijlstra <peterz@infradead.org>
To:     Mike Galbraith <efault@gmx.de>
Cc:     Mel Gorman <mgorman@techsingularity.net>,
        Ingo Molnar <mingo@kernel.org>,
        Vincent Guittot <vincent.guittot@linaro.org>,
        Valentin Schneider <valentin.schneider@arm.com>,
        Aubrey Li <aubrey.li@linux.intel.com>,
        Barry Song <song.bao.hua@hisilicon.com>,
        Srikar Dronamraju <srikar@linux.vnet.ibm.com>,
        LKML <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 2/2] sched/fair: Scale wakeup granularity relative to
 nr_running
Message-ID: <YVwZx1Mc6FQdWY6J@hirez.programming.kicks-ass.net>
References: <20210920142614.4891-1-mgorman@techsingularity.net>
 <20210920142614.4891-3-mgorman@techsingularity.net>
 <22e7133d674b82853a5ee64d3f5fc6b35a8e18d6.camel@gmx.de>
 <20210921103621.GM3959@techsingularity.net>
 <0cf76bb7701d4a37367773881c7d7c7bfceb455e.camel@gmx.de>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <0cf76bb7701d4a37367773881c7d7c7bfceb455e.camel@gmx.de>
Precedence: bulk

On Tue, Sep 21, 2021 at 02:32:54PM +0200, Mike Galbraith wrote:
> On Tue, 2021-09-21 at 11:36 +0100, Mel Gorman wrote:

> > FAIR_SLEEPERS primarily affects tasks that just became runnable and the
> > new task is trying to fit in without causing too much disruption based
> > on sysctl_sched_latency.
> 
> No, fair sleepers is all about sleeper wakeup preemption, I think
> you're thinking of fork initial placement.

Butting in in the middle of the thread (and I know there's still lots to
read)...

So the FAIR_SLEEPERS thing is about giving tasks that have slept a while
some extra credit to run sooner.

The classical example has always been a task that run 50% combined with
a task that runs 100%, what's fair? a 1:2 or 1:1 ratio? Strict fair
(runnable) scheduling will get you the 1:2, while intuitively having two
tasks with 100% combined CPU utilization 1:1 would be 'fair'.

FAIR_SLEEPERS gets you towards that 1:1, *provided* the period of that
50% is near sched_latency/2.

Another important factor for wakeup preemption has always been desktop
usage; can you still get responsive terminals while building a kernel,
how does firefox scroll during a kernel build etc..

(fwiw, firefox should start scrolling responsively and then bog down if
you keep on scrolling because it becomes a hog and has exhausted the
inital boost)

Also, I think the ChromeOS people have interactivity measures these
days.

All our traditinoal benchmarks miss out here; they're mostly throughput
oriented, and it is really easy to totally wreck interactivity while
getting great througput :/