Joe Mario

I just tested this patch on a single socket Sapphire Rapids system, running a 6.3.0-0.rc4.35.eln126.x86_64 kernel. I did see power savings at idle, (12% when C1 was not pinned, and...

To add a clarification to my earlier comment, there was no power savings when the cpus were busy. Those 8% and 12% power savings were only for idle systems.

> I did not thought this through completely yet, but I was thinking to have the "knobs" in a new configuration file under `/etc/tuned/something.conf`. Please see `/etc/tuned/cpu-partitioning-variables.conf` for example. @yarda,...

Hi Jason: Try this: gcc -g false_sharing_example.c -pthread -lnuma -o false_sharing.exe perf c2c record -F ./false_sharing.exe 40 Then hit ctrlC after 2-3 seconds. Joe

Hi Jason: Can you attach your c2c report output file? Joe

Hi Jason: A few more thoughts to consider. 14K samples for just that program or for just user space (--all-user) is nothing to worry about. That is sufficient. Remember, C2C...

Thank you Jaroslav for the info on dynamic tuning. I was unaware of those details.

Jan: Before you complete this work: 1) We need to have @prarit respond with confirmation that the algorithm he said TuneD needs to use is indeed valid. 2) The perf...

Hi Jan: > Joe wrote: > > The comments in this section of new code (lines 539-555) mention "energy_perf_bias". Unless I'm missing something, those comments should be mentioning energy_performance_preference. >...