Linux Lands Fix For Intel Hybrid CPU Frequency Scaling When Disabling E-Cores

[phoronix]

Phoronix: Linux Lands Fix For Intel Hybrid CPU Frequency Scaling When Disabling E-Cores

The Intel P-State CPU frequency scaling driver for the Linux kernel has received a fix to an issue that could lead to inadequate CPU frequency scaling behavior when running on a hybrid processor with E cores disable...

https://www.phoronix.com/news/Linux-P-State-Disable-E-Fix

[coder]

Ah, this is actually good for me. I don't have an affected CPU model yet, but I do some benchmarking work and was planning on disabling E-cores to measure P-only performance on Raptor Lake. For properly testing on E-only cores, I was planning to buy an Alder Lake-N board.

[abufrejoval]

Ah, this is actually good for me. I don't have an affected CPU model yet, but I do some benchmarking work and was planning on disabling E-cores to measure P-only performance on Raptor Lake. For properly testing on E-only cores, I was planning to buy an Alder Lake-N board.

I've used numactl to tie testing workloads to one set of cores or the other. It can't avoid kernel or other stuff running whereever they want to run, but the observed app at least will use whatever I tell it to use (unless it's doing its own numactl internally, which e.g. streess-ng will do).

I also play with set_pl to set power limits really low on my Alderlake board and gauge how the E-cores start actually outperforming the P-cores when Wattages force them near 1GHz speeds.

Just buying another system to test relative performance seems a bit much when you have both included and software to control them.

[coder]

Just buying another system to test relative performance seems a bit much when you have both included and software to control them.

Unless what I'm trying to measure also involves a lot of kernel time. Which it sometimes does.

Thanks for the suggestions/info. I did know about numactl, FWIW.