Date: Mon, 20 Mar 2017 13:50:09 +0100
From: Peter Zijlstra <peterz@infradead.org>
To: "Rafael J. Wysocki" <rjw@rjwysocki.net>
Cc: Linux PM <linux-pm@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>,
        Viresh Kumar <viresh.kumar@linaro.org>,
        Juri Lelli <juri.lelli@arm.com>,
        Vincent Guittot <vincent.guittot@linaro.org>,
        Patrick Bellasi <patrick.bellasi@arm.com>,
        Joel Fernandes <joelaf@google.com>,
        Morten Rasmussen <morten.rasmussen@arm.com>,
        Ingo Molnar <mingo@redhat.com>
Subject: Re: [RFC][PATCH 2/2] cpufreq: schedutil: Force max frequency on busy
 CPUs
Message-ID: <20170320125009.nmi3mvrxappjrvgo@hirez.programming.kicks-ass.net>
References: <4366682.tsferJN35u@aspire.rjw.lan>
 <2185243.flNrap3qq1@aspire.rjw.lan>
 <20170320103645.5jm2rwrx6bndwqmw@hirez.programming.kicks-ass.net>
 <1846666.Va7CsUJLIx@aspire.rjw.lan>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1846666.Va7CsUJLIx@aspire.rjw.lan>
User-Agent: NeoMutt/20170113 (1.7.2)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2471
Lines: 44

On Mon, Mar 20, 2017 at 01:35:12PM +0100, Rafael J. Wysocki wrote:
> On Monday, March 20, 2017 11:36:45 AM Peter Zijlstra wrote:
> > On Sun, Mar 19, 2017 at 02:34:32PM +0100, Rafael J. Wysocki wrote:
> > > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > > 
> > > The PELT metric used by the schedutil governor underestimates the
> > > CPU utilization in some cases.  The reason for that may be time spent
> > > in interrupt handlers and similar which is not accounted for by PELT.
> > > 
> > > That can be easily demonstrated by running kernel compilation on
> > > a Sandy Bridge Intel processor, running turbostat in parallel with
> > > it and looking at the values written to the MSR_IA32_PERF_CTL
> > > register.  Namely, the expected result would be that when all CPUs
> > > were 100% busy, all of them would be requested to run in the maximum
> > > P-state, but observation shows that this clearly isn't the case.
> > > The CPUs run in the maximum P-state for a while and then are
> > > requested to run slower and go back to the maximum P-state after
> > > a while again.  That causes the actual frequency of the processor to
> > > visibly oscillate below the sustainable maximum in a jittery fashion
> > > which clearly is not desirable.
> > > 
> > > To work around this issue use the observation that, from the
> > > schedutil governor's perspective, CPUs that are never idle should
> > > always run at the maximum frequency and make that happen.
> > > 
> > > To that end, add a counter of idle calls to struct sugov_cpu and
> > > modify cpuidle_idle_call() to increment that counter every time it
> > > is about to put the given CPU into an idle state.  Next, make the
> > > schedutil governor look at that counter for the current CPU every
> > > time before it is about to start heavy computations.  If the counter
> > > has not changed for over SUGOV_BUSY_THRESHOLD time (equal to 50 ms),
> > > the CPU has not been idle for at least that long and the governor
> > > will choose the maximum frequency for it without looking at the PELT
> > > metric at all.
> > 
> > Why the time limit?
> 
> One iteration appeared to be a bit too aggressive, but honestly I think
> I need to check again if this thing is regarded as viable at all.
> 

I don't hate the idea; if we don't hit idle; we shouldn't shift down. I
just wonder if we don't already keep a idle-seqcount somewhere; NOHZ and
RCU come to mind as things that might already use something like that.