linux-kernel - Re: [PATCH] thermal/governors: Remove integral

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:   Thu, 13 Oct 2022 16:58:10 +0100
From:   Lukasz Luba <lukasz.luba@....com>
To:     Qibo Huang <huangqibo.tech@...il.com>
Cc:     rafael@...nel.org, daniel.lezcano@...aro.org, amitk@...nel.org,
        linux-pm@...r.kernel.org, linux-kernel@...r.kernel.org,
        rui.zhang@...el.com
Subject: Re: [PATCH] thermal/governors: Remove integral_cutoff parameter, IPA
 is more regulated

Hi Qibo,

On 10/13/22 10:07, Qibo Huang wrote:
> Reason 1: If the integral_cutoff parameter is 0,
> the current temperature is greater than the control
> temperature, and err_integral continues to increase.
> If an abnormal situation occurs suddenly, the err_integral
> value will become very large. Even if the current
> temperature is lower than the control temperature,
> err_integral will always exist, which will cause the
> IPA to run out of control and cannot return to normal.
> This is the problem I'm actually having.
> 
> Reason 2: The integral_cutoff parameter is difficult to
> confirm the optimal, and can not bring good results.

Have you tired different 'k_i', 'k_po', 'k_pu' settings,
not the default from binding estimation?

These coefficients help to control the behavior of IPA
and how good the control loop could work.

The integral cutoff in default is set to 0, which means
we account the error when we overshoot the temperature
vs our desired control temperature value.

Please also pay attention to the type for those variables:
err, i, p and how they are treated (and when).
In current implementation we account 'err' negative, so we
reduce the power budget. Next time, even when we are below
control temperature, we still 'remember' those
overshoot mistakes, so we upfront shrink the possible
budget (to avoid overshooting). We clean the accumulated
error when the temperature drops below 'enable' trip point.

Could you describe a bit more what kind of problems you
observe, please?

Regards,
Lukasz