Date: Mon, 13 Mar 2006 01:46:05 -0800 (PST)
>From: Manu Ignatius <manuignatius@...>
>Subject: [oprof-cvs] time vs sample relation
> can anyone please tell me what is the relation between the
>oprofile samples and the time taken??? first i thought they
>are directly proportional, like when the no: of samples is
>large, the time taken also will be large, but in the tests
>conducted, i found out that the reverse is the case, Please
First, I suggest that in future you send usage questions to
oprofile-list@.... Many more community members
monitor that list versus the oprofile-commits mailing list.
As to your question, how do you know that "the reverse is the case"?
Assuming your analysis is correct, then the reason for these results
depends on what your profiling criteria is and other factors. Does your
platform support OProfile's event profiling and, if you built OProfile
from source, did you configure 'with-kernel-support'? In other words,
is OProfile using timer interrupt mode profiling or event-based
profiling? OProfile is a statistical profiler, so, while statistically
accurate, a given run may not be 100% accurate. And if you're using
timer-based profiling, you won't get any samples for interrupt handling
routines. If you're using event-based profiling, then specifiying
either cycles or instructions completed will give you the most accurate
"hot spot" profile, where number of samples would match up best with
actual time consumed by the target program. Specifying other types of
events will show you hot spots for those types of events, which may or
may not correlate to amount of time consumed by that section of code.
If you're analyzing at the instruction level, architectural "features"
can cause sample skew. See the chapter titled "Profiling interrupt
latency" in the user manual for some examples.
I hope this helps.
>Do You Yahoo!?
>Tired of spam? Yahoo! Mail has the best spam protection around