|
|
Subscribe / Log in / New account

Frequency-invariant utilization tracking for x86

Frequency-invariant utilization tracking for x86

Posted Apr 2, 2020 17:03 UTC (Thu) by jcm (subscriber, #18262)
Parent article: Frequency-invariant utilization tracking for x86

There is generally this notion that has existed for too long that frequency control is the only thing to be concerned about. In reality, there's a lot more an OS/platform can collaborate on, from frequency, to thermal pressure. On Arm servers, the CPPC extension to ACPI allows some of this to be enabled in a generic way and AMD extended this spec in their Zen2+ designs also. I would like to see generic CPPC/schedutil integration and adoption of such specs beyond just Arm platforms. To include thermal pressure and many other aspects as well. To that end, I'm planning for us to do some work on this and propose standard solutions in the months ahead. Very keen to hear from others who would like to discuss gaps in the current generic standards and how to plug those with a cross-industry solution.


to post comments

Frequency-invariant utilization tracking for x86

Posted Apr 3, 2020 23:33 UTC (Fri) by flussence (guest, #85566) [Link]

There was a patch[1] a few months ago, submitted by an AMD employee, trying to add CPPC2 awareness to the ondemand governor. It was rejected (rightly so IMO) for ignoring schedutil entirely and adding far too many knobs; the existing AMD CPB subdriver (which also ignores schedutil) only needs one boolean.

On some of the older CPUs I have there's also a 100mhzsteps feature flag in /proc/cpuinfo, I don't know if it's just not worth the effort or if nobody ever wrote the code to support it. The ACPI cpufreq driver definitely doesn't have 100MHz granularity.

[1]: https://lkml.org/lkml/2019/7/10/682


Copyright © 2025, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds