Elena Laskavaia
|
Re: new Application Profiler 2 and -lprofilingS
|
Elena Laskavaia
04/18/2008 10:17 AM
post7047
|
Re: new Application Profiler 2 and -lprofilingS
When using Sampling without any instrumentation overhead is neglectable.
When using Call Count instrumentation it is bigger. When profiling is
"stopped" it still collects data but does not send over the wire (but
collected data includes only count of functions calls, not timestamps).
When using Function Instrumentation every function has overhead on enter
and on exit, additionally when buffer is full the longest operation is
to dump butter to a file. This time (overhead) however is extracted from
final results. For multithreaded environment function instrumentation
won't work very well because it does not measure thread time, it measure
total time. So if two functions are run in parallel (pseudo parallel on
one processor) and finish at the same time, deep time would show them
taking same amount of physical time, even they took 2 times shorter. So
if you application has concurring threads you should ether use sampling
or system wide profiling.
Richard Virlouvet wrote:
>
> Dear Elena,
>
> I can see in the IDE4.5ApplicationProfilerTrial Wiki page, about
> application profiling, the following sentence: "It better works for
> one thread, because with many threads overhead of such measurement can
> change application behavior.".
>
> Is it the same for instrumented system profiling ?
>
> The overhead exist:
> - at any time ?
> - only when Tau is connected to the running process ?
> - or only when profiling ? (it means Tau is connected to the running
> process, but the new "start/stop" profiler feature is off)
>
> Thanks.
>
>
> _______________________________________________
> General
> http://community.qnx.com/sf/go/post7033
>
|
|
|