Intel P-State / CPUFreq CPU Scaling For Linux Gaming On Ubuntu 15.04

[phoronix]

Phoronix: Intel P-State / CPUFreq CPU Scaling For Linux Gaming On Ubuntu 15.04

It's been a while since last running any Intel P-State / CPUfreq scaling governor benchmarks on Phoronix. With a premium subscriber expressing interest in seeing a fresh comparison, here are some new numbers when running an Intel Core i7 Haswell CPU with NVIDIA GeForce graphics on Ubuntu 15.04 with the Linux 3.19 kernel and testing the different scaling drivers and governors.

http://www.phoronix.com/vr.php?view=22008

[schmidtbag]

I wonder if the drivers themselves have any impact on some of these results (particularly, the less consistent ones).

[Jeroen]

Thanks for this comparison Michael! However the other important test would be to see if there actually is any power saving, and if the CPUfreq Powersave performs considerably worse because it saves more power.

[Michael]

Thanks for this comparison Michael! However the other important test would be to see if there actually is any power saving, and if the CPUfreq Powersave performs considerably worse because it saves more power.

That's been covered in past comparisons... Generally, the P-State performance delivers the best performance-per-Watt.

[Gusar]

Thanks for this comparison Michael! However the other important test would be to see if there actually is any power saving, and if the CPUfreq Powersave performs considerably worse because it saves more power.

For such a test to really be interesting, you'd need a scenario that doesn't just stress the CPU/GPU full-time. Something like "open browser, visit phoronix.com, open music player and start a song, minimize browser and open photo viewer and wait for thumbnails to load, open a photo and apply a filter, close all apps", possibly with short breaks in between each step to simulate how you'd actually use the computer. This would need to be carefully scripted using tools that simulate mouse movement and button clicking, so it all happens exactly the same in each iteration. That would really tell you how the governors affect power consumption. Would be quite difficult to set up though.

[Guest]

I think there is a known bug with some powerclamp driver that sometimes goes full berzerk and eats away CPU even when idle.
People temporarily disabled pstate so the bug doesn't trigger, but when the module is blacklisted or not even compiled in, pstate generally works fairly well.
It also provides very good scaling and low scaling latencies (useful for gammers).

http://askubuntu.com/questions/482307/kidle-inject-uses-cpu-power-without-apparent-reason

https://www.kernel.org/doc/Documentation/thermal/intel_powerclamp.txt

[schmidtbag]

For such a test to really be interesting, you'd need a scenario that doesn't just stress the CPU/GPU full-time. Something like "open browser, visit phoronix.com, open music player and start a song, minimize browser and open photo viewer and wait for thumbnails to load, open a photo and apply a filter, close all apps", possibly with short breaks in between each step to simulate how you'd actually use the computer. This would need to be carefully scripted using tools that simulate mouse movement and button clicking, so it all happens exactly the same in each iteration. That would really tell you how the governors affect power consumption. Would be quite difficult to set up though.

That's really difficult to benchmark, or to make consistent. I actually would argue games are better since not all of them will max out a CPU core. On that note though, driver overhead could have an impact on the anomalies Michael found.

[Gusar]

That's really difficult to benchmark, or to make consistent.

I know, that's why I finished my post with "would be quite difficult to set up".

I actually would argue games are better since not all of them will max out a CPU core.

That's a good point.

I'm also thinking video playback (software decoded) would work well. Video has a fixed framerate, and specific steps the CPU needs to go through to decode the bitstream. Between those steps there's plenty of opportunity for the CPU to go idle. You control the software (a specific version of ffmpeg, for example) and the video clip, so such a test has easy reproduction.

[dimko]

Something like "open browser, visit phoronix.com, open music player and start a song, minimize browser and open photo viewer and wait for thumbnails to load, open a photo and apply a filter, close all apps", possibly with short breaks in between each step to simulate how you'd actually use the computer.

Please read the description of the article.

[Gusar]

Please read the description of the article.

The article measures the impact of governors. It's actual usage of a system the way I described it that is most affected by the governors! So what exactly should I read and what wisdom from it am I missing?