Monitoring CPU time

Sepherosa Ziehau sepherosa at gmail.com
Sun Sep 18 18:38:36 PDT 2016


Hi,

The bump upon each statclock is:
((cur_systimer - prev_systimer) * systimer_freq) >> 32

systimer_freq can be extracted from following sysctl in userspace:
sysctl kern.cputimer.freq

statclock is called at stathz frequency.

Thanks,
sephe


On Mon, Sep 19, 2016 at 12:46 AM, Stuart Nelson <stuartnelson3 at gmail.com> wrote:
> Hey all,
>
> I'm following up on some work from this old thread for monitoring CPU time:
> https://www.dragonflybsd.org/mailarchive/users/2010-04/msg00056.html
>
> The code I have is essentially the one shown in the link, but I'm attempting
> to find the actual number of seconds spent in each state. I'm doing this by
> dividing each value by clockrate.stathz, e.g.:
>
>
> user += cp_t[cpu].cp_user / clockrate.stathz;
>
>
> Relevant code is here:
> https://github.com/stuartnelson3/node_exporter/blob/2b5a581942ac31b501438d402274100df1f7d3d6/collector/cpu_dragonfly.go#L50-L98
>
> My question is about the units on struct members in kinfo_cputime (the
> source of cp_user et al.). The values I'm getting out seem to be growing at
> a rate that indicates I'm not looking at seconds, but something smaller.
>
> I'm looking at the rate of change of cpu time on my personal machine running
> dragonfly vs. a machine running linux. The implementation is the same: get
> user time, divide by 100Hz to get the value in seconds, and find the rate of
> change between two collections in fixed time window. The dragonfly rate of
> change seems to be larger by about 2 orders of magnitude, which is why I'm
> asking about the units.
>
> For reference, the dragonfly node I'm looking at is reporting ~200 increase
> per second in cpu time for user and sys with loadavg ~0.1%, whereas the
> linux node is reporting values <10 with loadavg ~15%.
>
> I'm improving dragonfly support for the node_exporter for Prometheus, a
> metrics and monitoring solution that is used mostly in the linux community.
> I'm assuming the linux implementation for finding cpu time in seconds is
> correct, and it's also the implementation used for finding cpu seconds for
> freebsd. It just struck me as unlikely that my old dell running dragonfly
> would have a rate of change at a fraction of the load that was so
> drastically different.
>
> If there is anything I can clarify don't hesitate to write back!
>
> Thanks,
> Stuart
>
>



-- 
Tomorrow Will Never Die



More information about the Users mailing list