CPU Utilisation Calculation

[rushi47]

Hey Tomas, Hope you are doing great :) I looked at your project and its awesome. I have few doubts, it will be great favour if you can help me with it. Specially interested in cpu utilisation part. Am trying to calculate the same, on my own. But it seems, whenever i calculate the value its not even close, can you verify approach once ?

I want to calculate cpu for specific time interval, just to give example. def something_get_execute() i want to calculate cpu usage for just this specific function.

For the same am doing following this, Getting values for user & system from here :

/sys/fs/cgroup/cpu,cpuacct/cpuacct.stat

Before function gets execute & after the function gets execute. After that i get delta for container by using below formula :

container_cpu_delta = (cpuacct.stat_final_user - cpuacct.stat_initial_user) / (cpuacct.stat_final_system - cpuacct_stat_initial_system)

Meanwhile i do the same, to get the value for host Before and After the function gets execute i get the value for User and system from :

/proc/stat

grep the first line and i get 1st & 3rd Param

cpu 205434 341 56826 3122952 48119 0 25640 0 0 0 i.e 205434 & 56826

And perform the same calculation host_cpu_delta = (proc_stat_final_user - proc_stat_init_user)/(proc_stat_final_system-proc_stat_init_user)

& Then finally i do

cpu_percentage = (container_cpu_delta/host_cpu_delta) * no_of_cores (am bit confused about this cores allocated to container or cores on the host)

However there are some observations:

• So many times i have seen the difference of (cpuacct.stat_final_user - cpuacct.stat_initial_user) i.e container is Greater than (proc_stat_final_user - proc_stat_init_user) of Host which i will feel little weird. As Host values contains the cycle of container as well
• Am confused about no of cores which to take while calculation, cpus allocated to container or all the cpus on host ?
• All things am doing from inside container, like reading all the values.

Please let me know if am missing anything in the calculation, it will be great help.

Thanks a lot :)

I created same issue Here also. Code am trying to work on is also in python. It will be great if you can comment on this.

[TomasTomecek]

I'd say that /proc/stat and cgroups use different type of values so that's why they are not consistent. So I guess you shouldn't mix them.

This is the code which performs the calculation:

https://github.com/TomasTomecek/sen/blob/master/sen/util.py#L176

I took it directly from docker's go code without thinking too much about it.

That's all I can help.

[rushi47]

oh ok Thanks for the answer @TomasTomecek

[rushi47]

If someone get stucked with same problem & want something like this : https://pypi.org/project/cgrputil/ , this is the library i created, feel free to use it :)