Re: default_idle time

[Piyush S. <sh...@cs...>]

Hi,

Thanks for your suggestions. We did our experiments once more with VTUNE
taking care of all the parameters and found out that both the profilers
give very similar results. This was very encouraging in our efforts to
move our setup to the open source profiler.

Thanks,
Piyush



On Wed, 6 Nov 2002, William Cohen wrote:

> Comparing the data between VTune and OProfile is a really good idea. See
> if the break down of the measurements are comparable. Is VTune shuffling
> the data to another machine? If so, what communication channel is it
> using to send the data? Maybe there is some side-effect from VTune's
> method of data collection. From what I understand VTune records a log of
> each sample. If the sample rate is high, then maybe there is some
> overhead associated with handling the samples. I haven't used VTune, so
> I don't know. What does it report when you collect data on an idle
> system? Does VTune show the system as idle?
>
> Are there any difference between the kernel being used for OProfile and
> VTune? Look at any differences between the experimental set up for
> OProfile and VTune.
>
> -Will
>
> Philippe Elie wrote:
> > Piyush Shivam wrote:
> >
> >>
> >>> 900 Mbps is only 112.5MB/s. The Intel PIII and P4 processors should have
> >>> little problem pushing 112MB/s onto the data bus. If the TCP/IP stack is
> >>> efficiently written (e.g. avoiding unneeded copies), then the processor
> >>> is not the limiting element. Other things such as the PCI bus (132MB/s)
> >>> or protocol handshaking could be limiting bandwidth. In these cases the
> >>> processor would be idle for a portition of the time. Why do you expect
> >>> the processor to be busier? Are you expecting to get 1GBps on the
> >>> channel and that the processor is the only possible bottle neck?
> >>
> >>
> >>
> >> Thanks for your response. We had done similar experiments with VTUNE and
> >> that shows the cpu being occupied for almost all the time with identical
> >> tests. We are trying to move our tests to oprofile because its an open
> >> source.
> >
> >
> > looking at top during an oprofiled run and vtun'ed runs
> > can give a clue where the cpu time is wasted, beside that
> > have you tried to check if when running oprofile you have
> > really 55 % cpu idle time. Run a long loop in C , $time a.out
> > on an idle system then running it with oprofile and your tcp
> > test, the real time given by the two run will give a good clue
> > if you have really 55% of idle time.I don't think than vtune
> > can waste55% of cpu time but your HW is a little what special
> > generating a lot of HW interrupts, iirc vtune work by remote
> > profiling, perhaps all these fact induce some artifact in
> > vtune measurements.
> >
> > Can you give more information on your system (SMP, kernel
> > version, what is roghly the interrupt rate of the tested
> > card)
> >
>
>