Intel Skylake CPUFreq vs. P-State Scaling Benchmarks On Linux 4.5

[phoronix]

Phoronix: Intel Skylake CPUFreq vs. P-State Scaling Benchmarks On Linux 4.5

For those curious about the performance impact between the CPUFreq and P-State scaling drivers and the different scaling governors when using an Intel Core i5 "Skylake" CPU with the latest Linux 4.5 kernel, here are some fresh benchmarks...

http://www.phoronix.com/scan.php?page=news_item&px=Intel-Linux-4.5-SKL-PState

[tessio]

Adding power usage along side with performance numbers would be great..

[Michael]

Adding power usage along side with performance numbers would be great..

FTA: "If you are curious about the performance-per-Watt or overall AC system power consumption differences, see some of our older results. I didn't do any power consumption results as this was just some quick weekend benchmarking for a one-page Phoronix article comparison (plus my lone USB WattsUp Pro power meter was busy with another system)."

[Kano]

Interesting to see such a gap between powersave and ondemand for GFX benchmarks. Maybe Intel should check that. 10 % is a bit too much.

[birdie]

CPUFreq powersave meanwhile yields very slow performance at the cost of power savings.

It actually works vs. PState Powersave which does NOT work.

Your graphs and your words, Michael, tell two different stories. It's like you're blind or something.

[Ropid]

It actually works vs. PState Powersave which does NOT work.

Your graphs and your words, Michael, tell two different stories. It's like you're blind or something.

That powersave governor has lower performance AND its power usage is a lot higher for that particular test. The kernel compilation takes a lot longer to complete and this then means terrible power efficiency compared to the other governors where the CPU can go back to sleep a lot earlier.

Check out that old article that was linked that has measurements of the power use of the machine. On page four and five of the article, there's similar results like here, with large differences between the governors. You can see that the Watts used are still pretty high despite the terrible performance. It's then overall just completely bad. You get both terrible performance and increased power usage.

[chuckula]

It actually works vs. PState Powersave which does NOT work.

Can you explain why it "does NOT work" exactly? Can you point to a specific bug report or other factual evidence on the LKML where Intel has said that Skylake's Pstate powersave is currently non-functional?

[Kano]

I guess he meant the cpufreq powersave, but this does something different than the rest, it keeps frequency low. The pstate powersave should work similar to the old ondemand variante. This seems to work, but if gpu load is there at the same time, something is different, maybe it checks a powerlimit that ondemand does not or whatever.

[birdie]

I meant that the PState governor does not actually respect the power saving mode, i.e. it doesn't stay at the lowest possible frequency/power consumption.

That powersave governor has lower performance AND its power usage is a lot higher for that particular test. The kernel compilation takes a lot longer to complete and this then

I've no idea where from you're pulling this conclusions and data.

[Ropid]

I meant that the PState governor does not actually respect the power saving mode, i.e. it doesn't stay at the lowest possible frequency/power consumption.

I've no idea where from you're pulling this conclusions and data.

What's happening is that choosing the lowest frequency is not the best to do for having the lowest power consumption. When running at a reduced frequency, completing the benchmark takes a lot longer than when running at full speed. When you then think about the power consumption over the whole benchmark, the CPU running at the highest frequency actually has lower power consumption for completing the task.

In that other article I linked to , you can see the power use between the reduced frequencies that cpufreq is using compared to the full frequency that P-State is choosing. You'll see that the power use doesn't drop as much as the performance drops.

Then, there's also idle. The power use when being idle drops a lot.

If you now put the two behaviors side-by-side for the same amount of time, you have a result where cpufreq lets the CPU run at the reduced frequency while the task takes longer to complete. Meanwhile, P-State runs the CPU at full speed for a certain portion of the time, then the task is complete, and the CPU goes to idle earlier.