2016-04-07

powerd++: Better CPU Clock Control for FreeBSD

Setting of P-States (power states a.k.a. steppings) on FreeBSD is managed by powerd(8). It has been with us since 2005, a time when the Pentium-M single-core architecture was the cutting edge choice for notebooks and dual-core just made its way to the desktop.

That is not to say that multi-core architectures were not considered when powerd was designed, but as the number of cores grows and hyper-threading has made its way onto notebook CPUs, powerd falls short.

Incentive

Don't you know it? You sit at your desk, reading technical documentation, occasionally scrolling or clicking on the next page link. The only (interactive) programs running are your web browser, an e-mail client and a couple of terminals waiting for input. There is a constant fan noise, which occasionally picks up for no apparent reason, making it a million times more annoying.

You can't work like this!

You start looking at the load, which is low but not minuscule. In the age of IMAP and node.js web browsers and e-mail clients are always a little busy. Still this is not enough to explain the fan noise.

You're running powerd to reduce your energy footprint (for various reasons), or are you? Yes you are. So you start monitoring dev.cpu.0.freq and it turns out your CPU clock is stuck at maximum like the speedometer of an adrenaline junkie with a death wish.

Something is wrong, your 15% to 30% load are way below the 50% default clock down threshold of powerd. You start digging, thinking you can tune powerd to do the right thing. Turns out you can't.

An Introduction to powerd

The following illustration shows powerd's operation on a dual-CPU system with two cores and hyper-threading each. That is not a realistic system today, but it saves space in the illustration and contains all the cases that need to be covered.

Note that …

  • … the sysctl(3) interface flattens the architecture of the CPUs into a list of pipelines, each presented as individual CPUs.
  • … powerd has the first CPU hard coded as the one controlling the clock frequency for all cores.
  • … powerd uses the sum of all loads to control the clock frequency.

powerd Architecture

Powerd using the sum of all loads to rate the overall load of the system allows single threaded loads to trigger higher P-States but comes at the cost of triggering high P-States with low distributed loads. The problem grows with the number of available cores. In the illustrated systems a mean load of 12.5% results in a 100% load rating. The same applies to a single quad-core CPU with hyper-threading.

Another problem resulting from this approach is that the optimal boundaries for the hysteresis changes with the number of cores. Also, to protect single core loads, powerd only permits boundaries from 0% to 100%. This results in powerd changing into the highest P-State at the drop of a needle and only clocking down if the load is close to 0.

The Design of powerd++

The powerd++ design has three significant differences. The way it manages the CPUs/cores/threads presented through the sysctl interface, the way that load is calculated and the way the target frequency is determined.

During its initialisation phase powerd++ assigns a frequency controlling core to each core, grouping them by the core that offers the handle to change the clock frequency. Unlike shown in the following illustration, all cores will always be controlled by dev.cpu.0, because the cpufreq(4) driver only supports global P-State changes. But powerd++ is built unaware of this limitation and will perform fine grained control the moment the driver offers it.

To rate the load within a core group, each core determines its own load and then passes it to the controlling core. The controlling core uses the maximum of the loads in the group as the group load. This approach allows single threaded applications to cause high load ratings (i.e. up to 100%), but having small loads on all cores in a group still results in a small load rating. Another advantage of this design is that load ratings always stay within the 0% to 100% range. Thus the same settings (including the defaults) work equally well for any number of cores.

Instead of using a hysteresis to decide whether the clock frequency should be increased, lowered or stay the same, powerd++ uses a target load to determine the frequency at which the current load would have rated as the target load. This approach results in quick frequency changes in either direction. E.g. given a target of 50% and a current load of 100% the new clock frequency would be twice the current frequency. To reduce sensitivity to signal noise more than two samples (5 by default) can be collected. This works as a low pass filter but is less damaging to the responsiveness of the system than increasing the polling interval.

powerd++ Architecture

Resources

The code is on github. A FreeBSD port is available as sysutils/powerdxx.

Afterthoughts

My experience in automotive and race car engineering came in handy. If your noise filter is not in O(1) (per frame), you're doing it wrong. If you have one control for many inputs a maximum or minimum are usually the right choice, the sum barely is. E.g. if you have 3 sensors that report 62°C, 74°C and 96°C, you want to adjust your coolant throughput to 96°C, not 232°C.

I hope that powerd++ will be widely used (within the community) and inspire the maintainers of cpufreq(4) to add support for per-CPU frequency controls.

TODOs

Currently the power source detection depends on ACPI, I need to implement something similar for older and non-x86/amd64 systems. Currently those just fall back to the unknown state.

No comments:

Post a Comment