Linux 5.11 Is Regressing Hard For AMD Performance With Schedutil

[phoronix]

Phoronix: Linux 5.11 Is Regressing Hard For AMD Performance With Schedutil

It's not the Grinch in 2020 that stole Christmas, but the Schedutil CPU frequency scaling governor on the in-development Linux 5.11 kernel that is thrashing performance for AMD Zen 2 and newer. Distributions like Ubuntu, Fedora, and Manjaro are beginning to use CPUFreq Schedutil by default on newer kernels and thus leading to a very bad initial/out-of-the-box experience with the current behavior on the early Linux 5.11 code.

http://www.phoronix.com/vr.php?view=29801

[aufkrawall]

Please always test with performance governors, unless you test governors themselves. Otherwise they might ruin every result...

[piotrj3]

Please always test with performance governors, unless you test governors themselves. Otherwise they might ruin every result...

I disagree, default test conditions should be default literally, unless you benchmark specific quirk like diffrent compiler or diffrent governor or something else.

[alberto-pv]

AMD should really pay you, Michael

[ThoreauHD]

They really should hire Michael, if they aren't going to stand up a proper kernel dev team. Better to beat down the nails with what you got, than get constantly blindsided by brokeness.

[cl333r]

AMD should really pay you, Michael

No, you're breathtaking!

[aufkrawall]

I disagree, default test conditions should be default literally, unless you benchmark specific quirk like diffrent compiler or diffrent governor or something else.

Why bother testing crap for 0.1W of power savings?
Those governors should be warned of and shunned, they severely hurt enduser Linux experience for often nonexistent power savings.

[zerothruster]

"The CPU power consumption of the EPYC 7702 was lower with Linux 5.11 Schedutil but on a performance-per-Watt basis was behind the Linux 5.10 efficiency."

Looks as if schedutil is really only concerned with minimising overall power consumption, which for data centres is perhaps a useful target, but otherwise only (sometimes) useful for laptops when they are running on battery power (or, perhaps, to stop them overheating which might be a consideration for intel hardware). No doubt developers with modern intel hardware can care about the reduced performance per watt (assuming it does also apply to intel and is not an "oh, we accidentally screwed up performance on amd, oh dear" case).

Michael, thanks for doing this. At least I can compile my own kernels. As you said, for distros the slowness will be a bad move if it doesn't get fixed.

[ssokolow]

Why bother testing crap for 0.1W of power savings?
Those governors should be warned of and shunned, they severely hurt enduser Linux experience for often nonexistent power savings.

That's why it's important to test them. To make sure bad defaults get called out.

[MadCatX]

Michael
Michael Larabel, saving the linux kernel performance once again! Companies whom you've saved millions of dollars in power expenses should pay you in gold. Anyway, enjoy my measly Paypal tip and merry xmas!