Date: Mon, 17 Jul 2017 21:23:09 +0200
From: Peter Zijlstra <peterz@infradead.org>
To: Andi Kleen <ak@linux.intel.com>
Cc: "Li, Aubrey" <aubrey.li@linux.intel.com>,
        Frederic Weisbecker <fweisbec@gmail.com>,
        Christoph Lameter <cl@linux.com>, Aubrey Li <aubrey.li@intel.com>,
        tglx@linutronix.de, len.brown@intel.com, rjw@rjwysocki.net,
        tim.c.chen@linux.intel.com, arjan@linux.intel.com,
        paulmck@linux.vnet.ibm.com, yang.zhang.wz@gmail.com, x86@kernel.org,
        linux-kernel@vger.kernel.org, daniel.lezcano@linaro.org
Subject: Re: [RFC PATCH v1 00/11] Create fast idle path for short idle periods
Message-ID: <20170717192309.ubn5muvc3u7htuaw@hirez.programming.kicks-ass.net>
References: <20170713083649.febfflfl5hafkko5@hirez.programming.kicks-ass.net>
 <16e12e23-6b28-f174-7c4b-4d719225cd3b@linux.intel.com>
 <20170713145311.z4zxlyd2dospeoqg@hirez.programming.kicks-ass.net>
 <4a577bd6-20b1-abb6-2153-f9870f0a721e@linux.intel.com>
 <20170713182820.sn3fjitnd3mca27p@hirez.programming.kicks-ass.net>
 <31170ac6-9db1-f0b8-4841-f1661c8ed6e1@linux.intel.com>
 <20170714153818.pjauqxebxyhs6ljp@hirez.programming.kicks-ass.net>
 <20170714155356.GH3441@tassilo.jf.intel.com>
 <20170714160648.tg2u6eo2id6gmnjz@hirez.programming.kicks-ass.net>
 <20170714162619.GJ3441@tassilo.jf.intel.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20170714162619.GJ3441@tassilo.jf.intel.com>
User-Agent: NeoMutt/20170609 (1.8.3)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1634
Lines: 40

On Fri, Jul 14, 2017 at 09:26:19AM -0700, Andi Kleen wrote:
> > And as said; Daniel has been working on a better predictor -- now he's
> > probably not used it on the network workload you're looking at, so that
> > might be something to consider.
> 
> Deriving a better idle predictor is a bit orthogonal to fast idle.

No. If you want a different C state selected we need to fix the current
C state selector. We're not going to tinker.

And the predictor is probably the most fundamental part of the whole C
state selection logic.

Now I think the problem is that the current predictor goes for an
average idle duration. This means that we, on average, get it wrong 50%
of the time. For performance that's bad.

If you want to improve the worst case, we need to consider a cumulative
distribution function, and staying with the Gaussian assumption already
present, that would mean using:

	 1              x - mu
CDF(x) = - [ 1 + erf(-------------) ]
	 2           sigma sqrt(2)

Where, per the normal convention mu is the average and sigma^2 the
variance. See also:

  https://en.wikipedia.org/wiki/Normal_distribution

We then solve CDF(x) = n% to find the x for which we get it wrong n% of
the time (IIRC something like: 'mu - 2sigma' ends up being 5% or so).

This conceptually gets us better exit latency for the cases where we got
it wrong before, and practically pushes down the estimate which gets us
C1 longer.

Of course, this all assumes a Gaussian distribution to begin with, if we
get bimodal (or worse) distributions we can still get it wrong. To fix
that, we'd need to do something better than what we currently have.