Re: [PATCH v2][RFC 1/2] Implement Ziegler-Nichols Heuristic

From: Daniel Lezcano
Date: Fri Feb 04 2022 - 06:50:36 EST


Thanks!

On 04/02/2022 12:48, Chetan Mistry wrote:
Hi Daniel,

Here is a link to the Compiled Jupyter Notebook with the performance analysis of the proposed system vs what is currently there:

https://nbviewer.org/gist/chemis01/0e86ad81508860659a57338dae8274f9 <https://nbviewer.org/gist/chemis01/0e86ad81508860659a57338dae8274f9>

Kind Regards,

Chetan

*From: *Lukasz Luba <lukasz.luba@xxxxxxx>
*Date: *Friday, 7 January 2022 at 13:39
*To: *Daniel Lezcano <daniel.lezcano@xxxxxxxxxx>
*Cc: *rafael@xxxxxxxxxx <rafael@xxxxxxxxxx>, linux-kernel@xxxxxxxxxxxxxxx <linux-kernel@xxxxxxxxxxxxxxx>, amitk@xxxxxxxxxx <amitk@xxxxxxxxxx>, Chetan Mistry <Chetan.Mistry@xxxxxxx>, rui.zhang@xxxxxxxxx <rui.zhang@xxxxxxxxx>, linux-pm@xxxxxxxxxxxxxxx <linux-pm@xxxxxxxxxxxxxxx>
*Subject: *Re: [PATCH v2][RFC 1/2] Implement Ziegler-Nichols Heuristic



On 1/7/22 11:52 AM, Daniel Lezcano wrote:
On 06/01/2022 14:16, Lukasz Luba wrote:

Thank you for fast response!

On 1/6/22 12:16 PM, Daniel Lezcano wrote:

Hi Lukasz,

On 06/01/2022 12:54, Lukasz Luba wrote:
Hi Daniel,

Could you have a look at this, please?

Yes, I had a quick look at the code and went to the algorithm
description.

Still digesting ...

On 12/17/21 6:49 PM, Chetankumar Mistry wrote:
Implement the Ziegler-Nichols Heuristic algorithm to better
estimate the PID Coefficients for a running platform.
The values are tuned to minimuse the amount of overshoot in
the temperature of the platform and subsequently minimise
the number of switches for cdev states.

Signed-off-by: Chetankumar Mistry <chetan.mistry@xxxxxxx>


This is the continuation of the previous idea to have
better k_* values. You might remember this conversation [1].

I've spent some time researching papers how and what can be done
in this field and if possible to plumb in to the kernel.
We had internal discussions (~2017) of one method fuzzy-logic that I
found back then, but died at the begging not fitting into this
IPA kernel specific environment and user-space. Your suggestion with
observing undershooting and overshooting results sparked better idea.
I thought it's worth to invest in it but I didn't have
time. We are lucky, Chetan was designated to help me and
experiment/implement/test these ideas and here is the patch set.

He's chosen the Ziegler-Nichols method, which shows really
good results in benchmarks (Geekbench and GFXbench on hikey960 Android).
The improved performance in Geekbench is ~10% (vs. old IPA).

+10% perf improvements sounds great. What about the temperature
mitigation (temp avg + stddev) ?

Chetan would respond about that with the link to the .html file.
We just have to create an official public server space for it.
I hope till Monday evening we would get something.


The main question from our side is the sysfs interface
which we could be used to trigger this algorithm for
better coefficients estimations.
We ask user to echo to some sysfs files in thermal zone
and start his/her workload. This new IPA 'learns' the system
utilization and reaction in temperature. After a few rounds,
we get better fitted coefficients.
If you need more background about the code or mechanisms, or tests,
I'm sure Chetan is happy to provide you those.

I'm worried about the complexity of the algorithm and the overhead
implied.

The k_* factors are tied with the system and the thermal setup (fan,
heatsink, processor, opp, ...). So IIUC when the factors are found, they
should not change and could be part of the system setup.

True, they are found and will be fixed for that board.


Would the algorithm fit better in a separate userspace kernel tooling?
So we can run it once and find the k_* for a board.

We wanted to be part of IPA in kernel because:
- the logic needs access to internals of IPA
- it would be easy accessible for all distros out-of-box
- no additional maintenance and keeping in sync two codes, especially
    those in some packages for user-space

Sorry, I'm not convinced :/

AFAICT, the temperature and the sampling rate should be enough
information to find out the k_*

We are allowing to overshoot the temperature by not capping the
power actors, but we have a safety net to not overshoot too much.
It's internal decision insdie IPA. Userspace would have to
re-implement whole IPA logic and take control over cooling
device states - which would contradict the decision from IPA
controlling the same thermal zone. The finding of coefficients
is by testing many values while running. The post processing
of the data (temp., power requests, frequency, etc) won't tell us
the the limits. We have to check them. That means the user-space tool
would have to re-implement major part of IPA, but also somehow
get the CPUs utilization, then use the obsolete user-space
governor API to experiment with them.


IMO, an userspace tool in ./tools/thermal/ipa is the right place

So if you give the tooling to the SoC vendor via the thermal ones, with
a file containing the temp + timestamp, they should be able to find the
k_* and setup their boards.

This feature is aimed for every user of the device, even w/o
expert knowledge in thermal/power. If vendor or OEM didn't
support properly the board, users could do this and they don't
have to be restricted (IMO).
Apart from that, I see vendors are rather interested in investing
in their proprietary solutions, not willing to share know-how
in open source power/thermal mechanisms. Then user is restricted
IMO in using the board. This might block e.g. research of some
PhD students, who have good ideas, but the restrictions of the platform
prevent them or cause the behavior of the board that they cannot
control. I would like to enable everyone to use fully the potential
of the HW/SW - even end-user.


Actually my opinion is the kernel should not handle everything and the
SoC vendor should at least do some work to setup their system. If they
are able to find out the sustainable power, they can do the same for the
right coefficients.

Ideally, yes, I would also like to see that. As you said a few months
ago during review of my former patches, a lot of this 'sustainable
power' entries in DT are Linaro contribution. I wish vendors also
contribute, but c'est la vie.


Additionally, the values can be stored in the Documentation for
different board and a documentation on how to use the tool.

Then up to the SoC vendor to setup the k_* in sysfs, so no need to
change any interface.

It wouldn't be for SoC vendor, but up to the OEM or board designer,
because the same SoC might have different thermal headroom thanks
to better cooling or bigger PCB, etc.

Right, s/SoC vendor/SoC platform/

I agree that these optimized k_* values might be shared in the kernel.
Ideally I would see them in the board's DT file, in the thermal zone,
but I'm afraid they are not 'Device description' so don't fit into DT
scope. They are rather optimizations of pure kernel mechanism.

where would be a good place for it? Maybe a new IPA Documentation/
sub-directory?

You can improve the documentation in:

Documentation/driver-api/thermal/power_allocator.rst

And if we agree on a tools/thermal/ipa, the documentation with examples
and some SoC reference can be put there also

If you are interested in those analyses we can find a way to share a>
.html file with the results from LISA notebook.

Yes,

Sure thing, let me arrange that.



IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.


--
<http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs

Follow Linaro: <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog