RE: [RFC/RFT][PATCH] cpufreq: intel_pstate: Work in passive mode with HWP enabled

["Doug Smythies" <dsmythies@telus.net>]

On 2020.05.26 01:19 Rafael J. Wysocki wrote:
>  to On Mon, May 25, 2020 at 10:57 PM Francisco Jerez
> > "Rafael J. Wysocki" <rafael@kernel.org> writes:
> > > On Mon, May 25, 2020 at 3:39 AM Francisco Jerez
> >
> > Why not HWP_MIN_PERF?  That would leave the HWP quite some room for
> > maneuvering (the whole HWP_MIN_PERF-HWP_MAX_PERF P-state range, it's not
> > like P-state selection would be entirely driven by the kernel), while
> > avoiding every failure scenario I was describing in my previous reply.

I have re-done my tests.
The problem that I observe seems specific to hwp itself
and not this patch and it's use in passive mode.
I see the exact same thing with intel_pstate/powersave.
[1] detail A.


Test: still simple single threaded load sweep,
at 347 hertz work/sleep frequency.
What do I see?

Unexpected frequency drops at around 70% load.
Example, from trace:

First, the thing has been going for awhile at 4.8 GHz.

Old epp ; new epp ; freq GHz; load % ; duration mS
80	  ; 82      ; 4.57    ; 61.94  ; 20.001
82      ; 80	; 4.57    ; 62.47  ; 40.003
80      ; 44      ; 3.73    ;	68.63  ; 62.009  <<<< What? Why freq down? Why long duration?
44      ;  0      ; 1.96    ; 100.23 ; 19.995  <<<< Even lower freq. load overruns.
 0      ; 73      ; 4.56    ; 82.93  ; 40.07   <<<< O.K. recovered, but damage done.
73      ; 46      ; 2.36    ;	79.19  ; 20.94   <<< now things oscillate a little.
46      ; 0       ; 1.9884  ;	100.24 ; 20.99
 0      ; 75      ; 4.5624  ;	82.1   ; 41.002  <<< Event ends. Next event in 487 milliseconds.

Observation: Events are often, but not always, preceded by a longer than normal duration.
However, long durations are also not out of the ordinary in passive mode.

And yes, the above trace was with DELAY_HWP 20,000, but I do have trace examples
with it at 5,000. This was just a particularly good example.

Observation (from looking at a lot of trace data): There are phase delays
between the two systems, intel_cpufreq and hwp, and sometimes they seem to
oscillate a little and fight each other. There maybe some problematic
work/sleep frequencies where the oscillation builds into a full blown
resonance. 
 
Why does hwp drop the frequency?

This system is otherwise fairly idle,
so maybe because the pll drops down during the non work periods.

Maybe HWP thinks the system is idle and drops the frequency.
I can eliminate the overruns by disabling deep idle states such
that the PLL vote is never relinquished, but it's not a fair test.

Note that the above response can be "tuned".
If we take the conversation algorithm from target frequency to EPP
and introduce and offset, the above can be improved.

At what cost? More sluggishness, for a large positive offset.
So, the overruns just move from the steady state side of the task to
when the task starts. I did not find if there is a "sweet spot"
between offset and system response, and I do not think there is value
added in trying.

Note: With original settings, I rarely observe a problem with the step
function response to a new task.

> 
> Actually, I have been thinking about the HWP min as an alternative
> that may be worth evaluating.
> 
> However, I would rather set the HWP min to something like 80% if the
> cpufreq request.

Yes, this is a good idea and should not suffer from the two servo systems
fighting each other.

I got 0 overruns, verses 2240 overruns with no min limitation (100 second test).

As for INTEL_CPUFREQ_TRANSITION_DELAY_HWP, I'll probably use
10 milliseconds moving forward, because that is what I am most
familiar with from years ago work on the this driver. But, I did
not observe any issue with 5 milliseconds.

[1] http://www.smythies.com/~doug/linux/intel_pstate/passive-hwp/passive-hwp-but-active-powersave.png

Other replaced graphs:

http://www.smythies.com/~doug/linux/intel_pstate/passive-hwp/passive-hwp-ondemand.png
http://www.smythies.com/~doug/linux/intel_pstate/passive-hwp/passive-hwp-schedutil.png
 
... Doug