[dtrace-discuss] Some of my nanoseconds are missing...

[Jim Mauro]

So, If I'm reading this right, you're profiling something that takes 
3,255 seconds
(54 minutes) - and you're worried about 3 microseconds?

It's noise. Forget it. The 3 microseconds you can not account for represents
.0001 percent (if I'm doing that math right) of the time window you're
profiling. There's nothing there. Move along folks, nothing to see here.

For measurements at this granularity, there's always going to be a fudge
factor of at least a few percent in terms of the accuracy of the times.

My apologies if I'm reading the numbers wrong...

/jim


Nik Clayton wrote:
> ... almost 3,000,000,000,000 of them.
>
> I'm instrumenting Sendmail (as some of you are probably bored of 
> hearing).
>
> Right now, I'm looking at where Sendmail spends its time when it's 
> creating queue files.
>
> I've discovered that a lot of time is being spent in calls to fdsync().
> The set of numbers I'm using shows 3,255,859,582,382 nanoseconds being 
> spent in fdsync() for a given workload.
>
> So I set about instrumenting fdsync(), and counting the time consumed 
> by  the kernel calls it makes under the same workload.
>
> Those figures show 286,960,093,086 nanoseconds taken up by kernel calls.
>
> So where did the other 3,000... go?  And can I get DTrace to 
> (automatically?) show me where the time is being spent?  I know I can 
> use the lockstat provider to see how long I'm spending waiting for 
> locks, but presumably, since those locks are being waited for by 
> kernel functions I've already accounted for that time.
>
> Ditto for the io provider, right?  Or do I have to instrument those too?
>
> More details, including pretty charts, at
>
> http://jc.ngo.org.uk/blog/2006/08/08/day-30-of-60-what-are-the-single-queue-directory-bottlenecks-pt-2/ 
>
>
> N
>
> PS:  Any chance D could grow a pragma to automatically insert commas 
> in large numbers?  It would make comparing results much easier.
> _______________________________________________
> dtrace-discuss mailing list
> dtrace-discuss at opensolaris.org