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:   Wed, 9 Jun 2021 09:50:00 +0000
From:   Quentin Perret <qperret@google.com>
To:     Peter Zijlstra <peterz@infradead.org>
Cc:     Beata Michalska <beata.michalska@arm.com>,
        Joel Fernandes <joel@joelfernandes.org>,
        Valentin Schneider <valentin.schneider@arm.com>,
        LKML <linux-kernel@vger.kernel.org>,
        Qais Yousef <qais.yousef@arm.com>,
        Vincent Guittot <vincent.guittot@linaro.org>,
        Viresh Kumar <viresh.kumar@linaro.org>
Subject: Re: iowait boost is broken
Message-ID: <YMCOyL8eiu9UpnEz@google.com>
References: <CAEXW_YTcO=hbmdq3nOx2RJfT2yPyoFnQx5niB38R2Lzpsp38bA@mail.gmail.com>
 <20210607191031.GA12489@e120325.cambridge.arm.com>
 <YL+tDv/EL5ogf/0w@hirez.programming.kicks-ass.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <YL+tDv/EL5ogf/0w@hirez.programming.kicks-ass.net>
Precedence: bulk

On Tuesday 08 Jun 2021 at 19:46:54 (+0200), Peter Zijlstra wrote:
> On Mon, Jun 07, 2021 at 08:10:32PM +0100, Beata Michalska wrote:
> > So back to the expectations.
> > The main problem, as I see it, is what do we actually want to achieve with
> > the I/O boosting? Is it supposed to compensate the time lost while waiting
> > for the I/O request to be completed or is is supposed to optimize the rate
> > at which I/O requests are being made. 
> 
> The latter, you want to increase the race of submission.
> 
> > Do we want to boost I/O bound tasks by
> > default, no limits applied  or should we care about balancing performance
> > vs power ? And unless those expectations are clearly stated, we might not
> > get too far with any changes, really.
> 
> You want to not increase power beyond what is needed to match the rate
> of processing I suppose.

Note that in some cases we also don't care about throughput, and would
prefer to keep the frequency for some unimportant IO bound tasks (e.g.
background logging deamons and such). Uclamp.max indicates this to some
extent.

Android for instance has a pretty good idea of the tasks it cares about,
and it'd be good to have a way to enable IO wait boosting only for
those.

Thanks,
Quentin