Re: How to find out which process cause a strange increasing cpu load on a certain core?

Hi Steven,

Thanks a lot for your reply.

It’s higher about 16% usr, 34% sys in top output but as I observe
pidstat has no any process consume this high cpu usage. After reboot
it can happen on a certain core on a certain VM, not fixed to any
core.

For ftrace, I mean function profile tracer by
echo 1 > /sys/kernel/debug/tracing/function_profile_enabled

I will try your suggestion.

Thanks again,
Brs,
Naruto

On Tue, 18 Jun 2019 at 20:02, Steven Rostedt <[email protected]> wrote:
>
> On Tue, 18 Jun 2019 18:40:13 +0700
> Naruto Nguyen <[email protected]> wrote:
>
> > Hello everyone,
> >
> > I have a kvm hypervisor and setup some VMs running OpenSUSE kernel
> > 4.0. After some reboot, I saw that sometime there is a VM that a
> > certain core has higher %usr and %sys than other cores. The system is
> > in ide state. I have some applications running on this VM, but there
> > is no issue in other VMs.
>
> How much higher? And is it always an issue with the same core on the
> same VM?
>
> >
> > If I enable ftrace in kernel or load any debug jprobe, the load is
>
> I'm assuming by "ftrace" you mean the full function tracer?
>
>  # echo function > /sys/kernel/tracing/current_tracer
>
> > back to normal. The pidstat command output does not show the high load
> > on one core like the "top" command. Could you please let me know if it
> > is the problem of "top' report higher usage? Any commands that can
> > help to troubleshoot this issue? I tried "perf" but most of the time
> > the perf output does not show any function that cause the increase as
> > well? Not sure, if the malfunction of top or anything else can add
> > more work load?
>
> I would suggest just tracing sched switches and interrupts. That
> shouldn't affect the load, as it is very low weight.
>
>  # cd /sys/kernel/tracing
>  # echo 1 > events/sched/sched_waking/enable
>  # echo 1 > events/sched/sched_switch/enable
>  # echo 1 > events/irq/enable
>
> And then monitor that.
>
> -- Steve