Performance of the scripts/cpufreqctl.sh script

[apanloco]

I don't have an issue running auto-cpufreq, I just wanted to share some information with benchmarks and script modification example.

When fiddling with auto-cpufreq, powertop and auditd to locate processes that wake the cpu up, ironically auto-cpufreq stands out on an otherwise completely idle system. Due to heavy usage of "expr" and "cat" the script functions run for a long time, and due to the nature of auto-cpufreq these functions are run periodically.

I made a proof of concept rewrite of scripts/cpufreqctl.sh/get_governor(), and these are the hyperfine benchmark results:

~ 6s ❯ hyperfine './get_governor_old.sh' './get_governor_new.sh'
Benchmark 1: ./get_governor_old.sh
  Time (mean ± σ):      12.8 ms ±   0.5 ms    [User: 10.2 ms, System: 2.1 ms]
  Range (min … max):    11.2 ms …  14.3 ms    226 runs

Benchmark 2: ./get_governor_new.sh
  Time (mean ± σ):       0.7 ms ±   0.2 ms    [User: 0.6 ms, System: 0.1 ms]
  Range (min … max):     0.5 ms …   1.8 ms    2653 runs

Summary
  ./get_governor_new.sh ran
   18.06 ± 4.08 times faster than ./get_governor_old.sh
~ ❯

Basically what this says it that on my system (Ubuntu 23.10, ThinkPad X1) just calling the original get_governor() takes 12ms and the new one takes 1ms.

The cpus are now working a lot less! Sleeping a lot more!

This is also confirmed just by running them (and we can see they have the same output):

~ ❯ time ./get_governor_old.sh
performance performance performance performance performance performance performance performance performance performance performance performance
./get_governor_old.sh  0,01s user 0,00s system 95% cpu 0,015 total

~ ❯ time ./get_governor_new.sh
performance performance performance performance performance performance performance performance performance performance performance performance
./get_governor_new.sh  0,00s user 0,00s system 85% cpu 0,001 total
~ ❯

15ms vs 1ms.

Old vs New implementation:

~ ❯ cat get_governor_old.sh
#!/bin/bash

function get_governor () {
  if [ -z $CORE ]
  then
    i=0
    ag=''
    while [ $i -ne $cpucount ]
    do
      if [ $i = 0 ]
      then
        ag=`cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor`
      else
        ag=$ag' '`cat /sys/devices/system/cpu/cpu$i/cpufreq/scaling_governor`
      fi
      i=`expr $i + 1`
    done
    echo $ag
  else
    cat /sys/devices/system/cpu/cpu$CORE/cpufreq/scaling_governor
  fi
}

get_governor

~ ❯ cat get_governor_new.sh
#!/bin/bash
function get_governor() {
  if [ -z "$CORE" ]; then
    ag=''
    for ((i=0; i<cpucount; i++)); do
      read -r current_governor < "/sys/devices/system/cpu/cpu$i/cpufreq/scaling_governor"
      ag="${ag}${current_governor} "
    done
    echo "${ag::-1}"
  else
    read -r current_governor < "/sys/devices/system/cpu/cpu$CORE/cpufreq/scaling_governor"
    echo "$current_governor"
  fi
}

get_governor

In the old implementation both cat and expr result in fork + exec. Eliminating that makes the scripts super fast. This is just a proof of concept for this particular function, and the sripts/cpufreqctl.sh script has many functions that can be optimized.

Hope you find it entertaining.