On Nov 7, 2007 6:22 PM, Roman Shaposhnik <[EMAIL PROTECTED]> wrote:
>   I didn't imply you did. I felt confused about your two statements
> and offered you an interpretation of them that seemed to fit hoping
> that it would be a useful first step to understanding what you
> actually had in mind.

OK :)

>   Do you have proof that the overhead is as great as you make it look?
> pthread_* operations are not exactly cheap. Even in userspace. Besides,
> the lock duration time is usually a relative value anyway.

Agreed.

We don't have Neelam's code so we don't know exactly what it does. So
there's too much guess work.

> I ran an
> experiment once using rdtsc assembly instruction for manually timing
> some mid-frequency user-space operations with and without DTrace
> enabled.
> The difference was there (and I shouldn't say it was
> negligible) but it wasn't all that impressive. I'd be curious to hear
> your side of the story.

I believe the overhead of firing a probe from userspace to the DTrace
framework is in the order of thousands of cycles. But I don't have
access to a Solaris system right now so I don't know. Whether DTrace
is good enough or not all depends on how long the lock is held. I am
interested in your experiment as well...

Rayson


>
> Thanks,
> Roman.
>
> P.S. As for the tools applicable to the problem at hand one can always
> use Sun Studio Performance Analyzer if small dilation is of a top
> concern and all that's needed is user-space application analysis.
>
>
_______________________________________________
dtrace-discuss mailing list
[email protected]

Reply via email to