Inconsistent values in /proc/stat

Kjeld Flarup · 07-31-2019, 08:24 AM

I'm looking for an explanation to how /proc/stat reports CPU usage. It is the basis for ex. the top command.

I have a 2 core ARMv7 processor running Linux kernel 4.14.34 with RT preempt patches.

It s my understanding that /proc/stat is counting ticks. My kernel is running 100 HZ, thus there would be 100 jiffies/s
Now I have made an observation on CPU1 that /proc/stat it may count more than 100 jiffies during a second. Usually the value is less than 100, also on a PC with Ubuntu.

This script shows how I do the calculation

Code:

b=`awk /cpu1/'{print $2+$3+$4+$5+$6+$7+$8+$9+$10+$11}' /proc/stat`
sleep 1
a=`awk /cpu1/'{print $2+$3+$4+$5+$6+$7+$8+$9+$10+$11}' /proc/stat`
expr $a - $b

Every time I boot I get a new value. It typically varies from 40 to 140 jiffies/s
And with every new value, busybox top applet show a new percentage of load.

Hope that somebody knows a good article on how this works.

greencedar · 08-02-2019, 08:10 AM

Does this article by github help your understanding?

https://github.com/Leo-G/DevopsWiki/...-is-calculated

If there article does not help, at the bottom of the article there are other references that may help your understanding.

Kjeld Flarup · 08-07-2019, 03:15 AM

Thanks for the link and references. Unfortunately they did not solve my problem, but made me sure that there is something wrong with figures I see in /proc/stat.

The unit should be in USER_HZ which since kernel 2.6 has been fixed to 100. Thus no more than 100 tick/s

However I can see that user space counts 100 ticks, thus the CPU is fully booked.
And then idle also counts 33 ticks on that CPU. Should not be possible.

ondoho · 08-08-2019, 02:56 AM

I suspect it's ARM-specific. Quick Search (tm).

Also look at this shell script:

Code:

#!/bin/bash 

# calculates cpu usage from values found in /proc/stat

delay=1 # 1 second between reads

color=''

# optional eyecandy
#~ tput civis
#~ trap 'tput cnorm' EXIT

# sleep as a builtin
for file in /usr/lib*/bash/sleep; do
    [ -r "$file" ] && enable -f "$file" sleep && break
done
# Portable enough?

while :; do
  # Get the first line with aggregate of all CPUs 
  read -r -a cpu_now </proc/stat
  # Get all columns but skip the first (which is the "cpu" string) 
  cpu_sum="${cpu_now[@]:1}" 
  # Replace the column seperator (space) with + 
  cpu_sum="${cpu_sum// /+}"
  # Get the delta between two reads 
  cpu_delta=$((cpu_sum - cpu_last_sum)) 
  # Get the idle time Delta 
  cpu_idle=$((cpu_now[4]- cpu_last[4])) 
  # Calc time spent working 
  cpu_used=$((cpu_delta - cpu_idle)) 
  # Calc percentage 
  cpu_usage=$((100 * cpu_used / cpu_delta)) 

  # Keep this as last for our next read 
  cpu_last=("${cpu_now[@]}") 
  cpu_last_sum=$cpu_sum 

# optional eyecandy
#~   case $cpu_usage in
#~    [5-7]?) color='\e[33m' ;; # yellow
#~    [8-9]?|100) color='\e[31m' ;; # red
#~  esac

  printf "\r\E[0KCPU ${color}%02d%%\E[0m" $cpu_usage 
  
  # Wait a second before the next read 
  sleep "$delay"
done