Re: [gentoo-user] CPU clock frequency crashing.

[Michael <confabulate@kintzios.com>]

[-- Attachment #1: Type: text/plain, Size: 2596 bytes --]

On Monday 30 December 2024 21:51:29 Greenwich Mean Time Alan Grimes wrote:
> Hey, I'm having usability problems in applications as my CPU clock
> frequency crashes down to about 530 mhz... I need it lock in a floor
> frequency of 1.25 ghz... I used
> cpupower frequency-set -d 1.25GHz  which does lift it up to a usable
> frequency but a day later it crashes down to garbage again. The machine
> is capable of 4.5 ghz or better on 32 cores. (it's a $3,200 processor, I
> shouldn't have to experience anything resembling a slowdown while the
> load average is under about 128...)
> 
> Basically, if it's not about to catch fire, then I need it to honor that
> frequency floor.

You haven't mentioned which CPU you are using.  If we're talking about AMD, 
then without diving into the weeds of the amd_pstate settings to understand 
the combination and interactions of CPPC and Pstate with your chosen scaling 
governor, I wouldn't know why a day later your minimum frequency setting has 
been reset.  I'd guess your scaling governor took over, but someone more 
studied on this topic can chime in.

From what I recall the userspace governor allows root to set the frequency 
statically by using a sysfs file.  Other governors e.g. schedutil will scale 
down the frequency by sampling the current load and make suitable predictions 
on impending demand.

> The next question is why it's even trying to scale down that low when
> there are a bunch of threads running flat out trying to show me youtube
> ads and crap. =\

Probably because a single thread plus some GPU cycles will process the youtube 
ads adequately, while the rest of the cores are not lifting any weight at the 
time.  Setting a floor frequency for the CPU when it does not need it is both 
unnecessary and wasteful when any and all variations in frequency are being 
controlled as needed by the firmware.  When I emerge big packages with all 
cores set in MAKEOPTS, I can observe individual cores boosting selectively for 
a few seconds before they drop slightly back waiting for their turn to max out 
again.  I have noticed the same round-robin behaviour boosts CPU cores 
momentarily when running cpuburn and other load maximizing benchmarking 
software.  I understand this is how the chipset algo controls the thermal load 
of the CPU.  Some tweaking is offered by the BIOS firmware, but the general 
recommendation is to leave it alone to do its thing, unless you are trying to 
achieve some wild benchmarking record for bragging rights and the $3,200 price 
tag going up in smoke is not a significant consideration for you.

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 833 bytes --]