Accuracy/reliability of "top" CPU usage numbers

[walter harms]

Am 13.11.2014 16:17, schrieb Grant Edwards:
> How accurate/reliable are the cpu usage numbers shown by busybox "top"
> for long-running multi-threaded processes?
> 
> Are they simply extract and shown as-is from /proc/<pid>/<whatever>,
> or is the top code actually doing something?
> 
> If it's the former, this is probably off-topic for the busybox
> list...
> 
> I'm working on performance issues for a smallish daemon which uses a
> timerfd to trigger periodic polls of some custom hardware (the
> hardware is mapped into userspace), and when it sees changes in input
> values it distributes those new values to some client apps via Unix
> domain sockets.  [I've acually commented out the actual socket data
> read/write code, so that isn't even happening.] The polling thread
> runs with a high real time priority (Posix FIFO scheduler, priority=99).
> 
> I'm using busybox "top" (on an Atmel ATM9 w/ kernel 2.6.33.7 with
> PREEMPT enable) to see how much CPU the daemon is using.
> 
> What's puzzling me is that the CPU usage show by "top" varies wildly
> depending on the frequency at which set the timerfd for polling the
> hardware.  For example:
> 
>       Poll Time (ms)   CPU usage (%)
>           20.8              5
>           20.0             16
>           16.8              3          
>           14.8              4
>           12.0              3
>           10.8              4
>           10.0             20
>            9.6              3
>            8.8              3
>            6.8              4
>            5.0              5
>            4.0              5
> 

Hi grant,
the values for 10,20 look strange the rest is more or less in line.
can you please check von 30,40 also ? (perhaps +-1)

IMHO the values of top are an estimate and it is highly recommended
to do something like you have done a take a time series to find a trend
instead of a "real" value.

re,
 wh


> These CPU usage numbers are pretty stable and repeatable (they don't
> vary over time or between runs by more than 1).
> 
> Is my app really using 6X as much CPU when polling at 10.0ms as it is
> when polling at 10.8ms or 9.6ms?  Or am I running into some sort of
> issue with sampling/aliasing and they way CPU usage is measured?
>