Fwd: kernel function probe overhead

[Frank Ch. Eigler]

aubreylee wrote;

> Thanks Cody.
> Here is what I got:
>
> ============================================================
> $ sudo stap -t test.stp
> -----------------------------------------
> [...]
> kernel.function("arch_cpu_idle_enter@arch/x86/kernel/process.c:277").call,
> (./test.stp:14:1), hits: 2962, cycles:
>
>>400min/30951avg/233043max, variance: 824817112, from: kernel.function("arch_cpu_idle_enter@arch/x86/kernel/process.c:277").call
>
> May I know what does variance value mean here?

This is normal statistical variance, i.e., the square of the standard
deviation.  That means that the bulk of the probe execution times were
within sqrt(824817712)=28719 cycles from the mean, if the distribution
were roughly normal.


> ----- refresh report:
>>'__global_t_start' lock contention occurred 2853 times
>
> Would it be nice to report time cost of lock contention?

That would be roughly #-contentions * TRYLOCKDELAY (default 10us),
so about 29 ms.


> Anyway, instead of reporting lock contention, it would be nice if we
> can reduce it, :)

Sure, but a good first step is to know the magnitude of the problem.


- FChE