Message-ID: <20201202155452.GK3021@hirez.programming.kicks-ass.net>
Date:   Wed, 2 Dec 2020 16:54:52 +0100
From:   Peter Zijlstra <peterz@...radead.org>
To:     Mel Gorman <mgorman@...e.de>
Cc:     "Rafael J. Wysocki" <rjw@...ysocki.net>,
        Ingo Molnar <mingo@...nel.org>,
        Thomas Gleixner <tglx@...utronix.de>,
        Vincent Guittot <vincent.guittot@...aro.org>,
        Morten Rasmussen <morten.rasmussen@....com>,
        dietmar.eggemann@....com, patrick.bellasi@...bug.net,
        lenb@...nel.org, linux-kernel@...r.kernel.org,
        valentin.schneider@....com, ionela.voinescu@....com,
        qperret@...gle.com, viresh.kumar@...aro.org
Subject: Re: [RFC] Documentation/scheduler/schedutil.txt

On Wed, Dec 02, 2020 at 02:18:35PM +0000, Mel Gorman wrote:
> For whatever reason, this only arrived in my inbox today.

It's as if the gods of email knew you was busy before ;-)

> On Fri, Nov 20, 2020 at 08:55:27AM +0100, Peter Zijlstra wrote:
> > PELT (Per Entity Load Tracking)
> > -------------------------------
> > 
> > With PELT we track some metrics across the various entities, from individual
> 
> s/entities/scheduler entities/ so that someone will recognise
> sched_entity in the code.
> 
> > tasks to task-group slices to CPU runqueues. As the basis for this we use an
> > EWMA, each period (1024us) is decayed such that y^32 = 0.5. That is, the
> 
> Expand -- Exponentially Weighted Moving Average (EWMA). The cc list
> should recognise it automatically, but maybe not a first entrant to
> kernel/sched. kernel/sched is littered with institutional knowledge.

It is, we could do with moar comments.

> > most recent 32ms contribute half, while the rest of history contribute the
> > other half.
> > 
> 
> IIRC, this 32ms is tied to the value of LOAD_AVG_PERIOD and the length
> of the ewma_sum series below. Might be worth expanding a little further.

It is LOAD_AVG_PERIOD. Some people (re)generate the PELT tables with a
different period (16 and 64 are common).

Not sure what there is to expand; the whole of it is: y^32=0.5. We had
to pick some half-life period, 32 seemed like a good number.

> > r_het is determined as the average performance difference between a big and
> > LITTLE core when running at max frequency over 'relevant' benchmarks.
> > 
> 
> r_het is never mentioned again so it's not immediately obvious what it
> ties into. I assume het is short for heterogeneous and is simply another
> way of looking at current CPU compute power vs potential CPU compute power
> (be that due to DVFS or big.LITTLE).

It's now called r_cpu, but yes.

> > The result is that the above 'running' and 'runnable' metrics become invariant
> > of DVFS and Heterogenous state. IOW. we can transfer and compare them between
> > CPUs.

^^^ that is the consequence; being able to directly compare the numbers
across CPUs.

> > 
> > For more detail see:
> > 
> >  - kernel/sched/pelt.h:update_rq_clock_pelt()
> >  - arch/x86/kernel/smpboot.c:"APERF/MPERF frequency ratio computation."
> > 
> 
> The role and importance of frequency invariance is mentioned but it
> could be more explicit. However, looking at update_rq_clock_pelt may be
> enough of a clue. 
> 
> Either way, decoding this document fully will require someone to spend a
> lot of time on the source and then rereading this document. That's
> probably a good thing.

:-)

> > UTIL_EST / UTIL_EST_FASTUP
> > --------------------------
> > 
> > Because periodic tasks have their averages decayed while they sleep, even
> > though when running their expected utilization will be the same, they suffer a
> > (DVFS) ramp-up after they become runnable again.
> > 
> 
> s/they become runnable again/they are running again/ ? Maybe refer as
> "running" because it's only once they are on the CPU and running that
> DVFS comes into play?

OK.

> > To alleviate this (a default enabled option) UTIL_EST drives an (IIR) EWMA
> 
> Expand IIR -- Immediate Impulse Reponse?

Infinite Impuse Response

> > with the 'running' value on dequeue -- when it is highest. A further default
> > enabled option UTIL_EST_FASTUP modifies the IIR filter to instantly increase
> > and only decay on decrease.
> > 
> > A further runqueue wide sum (of runnable tasks) is maintained of:
> > 
> >   util_est := \Sum_t max( t_running, t_util_est_ewma )
> > 
> > For more detail see: kernel/sched/fair.h:util_est_dequeue()
> > 
> 
> It's less obvious what the consequence is unless the reader manages to
> tie the IO-wait comment in "Schedutil / DVFS" to this section.

I'm not entirely sure I follow. The purpose of UTIL_EST is to avoid
ramp-up issues and isn't related to IO-wait boosting.

> > Schedutil / DVFS
> > ----------------
> > 
> > Every time the scheduler load tracking is updated (task wakeup, task
> > migration, time progression) we call out to schedutil to update the hardware
> > DVFS state.
> > 
> > The basis is the CPU runqueue's 'running' metric, which per the above it is
> > the frequency invariant utilization estimate of the CPU. From this we compute
> > a desired frequency like:
> > 
> >              max( running, util_est );	if UTIL_EST
> >   u_cfs := { running;			otherwise
> > 
> >   u_clamp := clamp( u_cfs, u_min, u_max )
> > 
> >   u := u_cfs + u_rt + u_irq + u_dl;	[approx. see source for more detail]
> > 
> >   f_des := min( f_max, 1.25 u * f_max )
> > 
> > XXX IO-wait; when the update is due to a task wakeup from IO-completion we
> > boost 'u' above.
> > 
> > This frequency is then used to select a P-state/OPP or directly munged into a
> > CPPC style request to the hardware.
> > 
> > XXX: deadline tasks (Sporadic Task Model) allows us to calculate a hard f_min
> > required to satisfy the workload.
> > 
> > Because these callbacks are directly from the scheduler, the DVFS hardware
> > interaction should be 'fast' and non-blocking. Schedutil supports
> > rate-limiting DVFS requests for when hardware interaction is slow and
> > expensive, this reduces effectiveness.
> > 
> 
> Is it worth explicitly mentioning that a key advantage over
> hardware-based approaches is that schedutil carries utilisation state on
> CPU migration? You say that it is tracked but it's less obvious why that
> matters as a pure hardware based approach loses utilisation information
> about a task once it migrates.

Not sure that was the exact goal of the document; I set out to describe
schedutil.

> Even moving note 3 below into this section and expanding it with an
> example based on HWP would be helpful.

I might not be the best person to talk about HWP; even though I work for
Intel I know remarkably little of it. I don't even think I've got a
machine that has it on.

Latest version below... I'll probably send it as a patch soon and get it
merged. We can always muck with it more later.

---

NOTE; all this assumes a linear relation between frequency and work capacity,
we know this is flawed, but it is the best workable approximation.


PELT (Per Entity Load Tracking)
-------------------------------

With PELT we track some metrics across the various scheduler entities, from
individual tasks to task-group slices to CPU runqueues. As the basis for this
we use an Exponentially Weighted Moving Average (EWMA), each period (1024us)
is decayed such that y^32 = 0.5. That is, the most recent 32ms contribute
half, while the rest of history contribute the other half.

Specifically:

  ewma_sum(u) := u_0 + u_1*y + u_2*y^2 + ...

  ewma(u) = ewma_sum(u) / ewma_sum(1)

Since this is essentially a progression of an infinite geometric series, the
results are composable, that is ewma(A) + ewma(B) = ewma(A+B). This property
is key, since it gives the ability to recompose the averages when tasks move
around.

Note that blocked tasks still contribute to the aggregates (task-group slices
and CPU runqueues), which reflects their expected contribution when they
resume running.

Using this we track 2 key metrics: 'running' and 'runnable'. 'Running'
reflects the time an entity spends on the CPU, while 'runnable' reflects the
time an entity spends on the runqueue. When there is only a single task these
two metrics are the same, but once there is contention for the CPU 'running'
will decrease to reflect the fraction of time each task spends on the CPU
while 'runnable' will increase to reflect the amount of contention.

For more detail see: kernel/sched/pelt.c


Frequency- / CPU Invariance
---------------------------

Because consuming the CPU for 50% at 1GHz is not the same as consuming the CPU
for 50% at 2GHz, nor is running 50% on a LITTLE CPU the same as running 50% on
a big CPU, we allow architectures to scale the time delta with two ratios, one
Dynamic Voltage and Frequency Scaling (DVFS) ratio and one microarch ratio.

For simple DVFS architectures (where software is in full control) we trivially
compute the ratio as:

	    f_cur
  r_dvfs := -----
            f_max

For more dynamic systems where the hardware is in control of DVFS (Intel,
ARMv8.4-AMU) we use hardware counters to provide us this ratio. For Intel
specifically, we use:

	   APERF
  f_cur := ----- * P0
	   MPERF

	     4C-turbo;	if available and turbo enabled
  f_max := { 1C-turbo;	if turbo enabled
	     P0;	otherwise

                    f_cur
  r_dvfs := min( 1, ----- )
                    f_max

We pick 4C turbo over 1C turbo to make it slightly more sustainable.

r_cpu is determined as the ratio of highest performance level of the current
CPU vs the highest performance level of any other CPU in the system.

  r_tot = r_dvfs * r_cpu

The result is that the above 'running' and 'runnable' metrics become invariant
of DVFS and CPU type. IOW. we can transfer and compare them between CPUs.

For more detail see:

 - kernel/sched/pelt.h:update_rq_clock_pelt()
 - arch/x86/kernel/smpboot.c:"APERF/MPERF frequency ratio computation."
 - Documentation/scheduler/sched-capacity.rst:"1. CPU Capacity + 2. Task utilization"


UTIL_EST / UTIL_EST_FASTUP
--------------------------

Because periodic tasks have their averages decayed while they sleep, even
though when running their expected utilization will be the same, they suffer a
(DVFS) ramp-up after they are running again.

To alleviate this (a default enabled option) UTIL_EST drives an Infinite
Impulse Response (IIR) EWMA with the 'running' value on dequeue -- when it is
highest. A further default enabled option UTIL_EST_FASTUP modifies the IIR
filter to instantly increase and only decay on decrease.

A further runqueue wide sum (of runnable tasks) is maintained of:

  util_est := \Sum_t max( t_running, t_util_est_ewma )

For more detail see: kernel/sched/fair.c:util_est_dequeue()


UCLAMP
------

It is possible to set effective u_min and u_max clamps on each CFS or RT task;
the runqueue keeps an max aggregate of these clamps for all running tasks.

For more detail see: include/uapi/linux/sched/types.h


Schedutil / DVFS
----------------

Every time the scheduler load tracking is updated (task wakeup, task
migration, time progression) we call out to schedutil to update the hardware
DVFS state.

The basis is the CPU runqueue's 'running' metric, which per the above it is
the frequency invariant utilization estimate of the CPU. From this we compute
a desired frequency like:

             max( running, util_est );	if UTIL_EST
  u_cfs := { running;			otherwise

  u_clamp := clamp( u_cfs, u_min, u_max )

  u := u_cfs + u_rt + u_irq + u_dl;	[approx. see source for more detail]

  f_des := min( f_max, 1.25 u * f_max )

XXX IO-wait; when the update is due to a task wakeup from IO-completion we
boost 'u' above.

This frequency is then used to select a P-state/OPP or directly munged into a
CPPC style request to the hardware.

XXX: deadline tasks (Sporadic Task Model) allows us to calculate a hard f_min
required to satisfy the workload.

Because these callbacks are directly from the scheduler, the DVFS hardware
interaction should be 'fast' and non-blocking. Schedutil supports
rate-limiting DVFS requests for when hardware interaction is slow and
expensive, this reduces effectiveness.

For more information see: kernel/sched/cpufreq_schedutil.c


NOTES
-----

 - On low-load scenarios, where DVFS is most relevant, the 'running' numbers
   will closely reflect utilization.

 - In saturated scenarios task movement will cause some transient dips,
   suppose we have a CPU saturated with 4 tasks, then when we migrate a task
   to an idle CPU, the old CPU will have a 'running' value of 0.75 while the
   new CPU will gain 0.25. This is inevitable and time progression will
   correct this. XXX do we still guarantee f_max due to no idle-time?

 - Much of the above is about avoiding DVFS dips, and independent DVFS domains
   having to re-learn / ramp-up when load shifts.

Powered by blists - more mailing lists