A Fix Has Been Proposed For The Slower AMD Performance On Linux 5.11

[phoronix]

Phoronix: A Fix Has Been Proposed For The Slower AMD Performance On Linux 5.11

With the in-development Linux 5.11 kernel there are many great features and improvements especially for AMD users with some new drivers and other pleasant enhancements. But as I outlined back on Christmas day: Linux 5.11 Is Regressing Hard For AMD Performance With Schedutil. Fortunately, a fix is now en route to the Linux 5.11 kernel for fixing that performance regression affecting AMD Zen 2/3 desktops and servers...

http://www.phoronix.com/scan.php?page=news_item&px=Linux-5.11-AMD-Inv-Freq-Patch

[Linuxxx]

I think it's important to note that the developer himself calls the currently proposed solution a »stop-gap« one; hopefully a proper fix for this problem really makes it into Linux 5.12, as planned.
(Which, by the wording, would help Intel systems too, it seems, since Intel apparently is already making use of this approach to at least kinda solve schedutil's problems.)

Here's the relevant quote:

2. PROPOSED SOLUTION (raise freq_max in schedutil formula) The solution we implement here is a stop-gap one: when the driver is acpi_cpufreq and the machine an AMD EPYC, schedutil will use max_boost instead of max_P as the value for freq_max in its formula freq_next = 1.25 * freq_max * util essentially giving freq_next some more headroom to grow in the over-utilized case. This is the approach also followed by intel_pstate in passive mode. The correct way to attack this problem would be to have schedutil detect over-utilization and select freq_max irrespective of the util value, which has no meaning at that point. This approach is too risky for an -rc5 submission so we defer it to the next cycle.

[bezirg]

Great news, looking forward to the new benchmark results.

[milkylainen]

I solved that long time ago: Scratching power saving configuration shenanigans.

[Guest]

I just use ondemand or performance, but this is great news for people who use the default of schedutil.

[andrei_me]

Amazing patch, all formatted, explained and so on, seems like an academic paper

[QwertyChouskie]

I solved that long time ago: Scratching power saving configuration shenanigans.

Less power used means less heat output, which can then allow the processor to boost up more during heavy loads, resulting in actually better performance. Especially in systems with iGPUs, less CPU power used means more thermal headroom and power budget for running the GPU faster (which is usually the bottleneck in games).