To understand Helix_schedutil you must first understand the original schedutil algorithm.
Here it is:
next_freq = maxfreq + (maxfreq >> bitshift) * util/maxcapacity
The most obvious difference of this algorithm is that it moves away from the idea of scaling frequencies up or down which were used in previous generations of governors.
Instead the aim of the above algorithm is to calculate the most appropriate frequency for the TOTAL CPU load.
NOTE: This is TOTAL load on CPU, not just load for the current frequency step as Interactive used to calculate with.
Now, for you numberphiles like myself that like understanding algorithms... Let's break it down:
"util/maxcapacity = Load."
The above creates a percentage value in decimal format (80% = 0.8) which represents the TOTAL load on CPU.
the algorithm now reads the following way:
next_freq = maxfreq + (maxfreq >> bitshift) * load
"maxfreq + (maxfreq >> bitshift)"
Essentially the aim of the above is to ensure that next_freq is always a little higher than the exact value needed to cover the load.
Bitshift: (paraphrasing @ZeroInfinity
) in programming the ">>" mathematical function allows for shifting the binary values towards the direction of the arrows by "N" times.
In this case it is towards the right.
The relationship between "N" and the calculation in the "()" is as follows:
Bitshift = 1 = maxfreq/2
Bitshift = 2 = maxfreq/4
Bitshift = 3 = maxfreq/8
If the "+()" didn't exist in the algorithm, the chosen frequency would be exactly enough to cover the load.
If load is 0.6, aka 60%, all you need is a frequency = 60% of max frequency.
This would be bad since it doesn't leave any capacity/bandwidth leftover for inevitable bumps in load, nor space for EAS itself to run. Thus inevitably creating lags.
To keep a bit of free bandwidth you add "(maxfreq >> bitshift)".
Finally the problem I encountered, if bitshift = 2, then the result of the algorithm is that any load above 0.8 will result in a next_freq HIGHER than maxfreq. - This is your tipping point. As any load higher than 80% will wake up a new CPU.
Which means you have still about 20% of the CPU's max capacity being unused. Such a CPU is only 80% efficient.
Therefore by increasing bitshift to 3, the algorithm reads:
"maxfreq+(maxfreq/8)*load = next_freq"
This way you can use 89% of capacity before reaching max frequency of the CPU.
With bitshift=4 it reads:
"maxfreq+(maxfreq/16)*load = next_freq"
This allows you to use up to 94% total CPU load before reaching max frequency.
While this is great for improving efficiency at the higher frequencies, it doesn't leave enough bandwidth when calculating lower frequencies, and creates lag when load spikes at lower frequencies.
Update to the explanation:
After being inspired by the concept of @ZeroInfinity
's new governor - Energy-DCFC, I decided to carry out a couple of tests on HTC 10 using variations of Helix_Schedutil.
The focus was stress-testing by increasing the current frequency load above 100%. (AKA Use up all of the bandwidth of the current frequency step.)
After the testing me and Zero worked on this new version of Helix_Schedutil.
The current behaviour of the governor is the following:
- Boost frequencies when load is below Target_Load1. (Boost can be increased by DECREASING bit_shift1 value.)
- Between Target_Loads there is no bit_shift at all. The governor just uses the following algorithm instead - (max_freq*util/max = next_freq)
- Loads higher than Target_Load2 will be THROTTLED. Bit_Shift2 here is subtracted rather than added. (Throttle effect can be increased by DECREASING bit_shift2 value.)
The result is that low freqs have spare bandwidth to avoid lags, middle frequencies leave no extra bandwidth at all, while higher frequencies are throttled to save battery.
Another focus of the governor update is to reduce overhead as much as possible. This results in a very responsive governor which isn't overly demanding on battery life.
Schedtune.boost values recommended for use with this governor:
Energy-DCFC is still recommended for those who prefer battery life over performance, but if you prefer greater performance then this governor can be used without making you feel guilty about wasting battery.