On Jul 2, 2008, at 8:44 AM, richard pan wrote:
> Sorry to bothering you guys.
> I have two questions that's so hard for me to find the answers.
> Is there any references that I can use to find a formula to
> calculate overhead memory and the difference performance when
> testing using different number of processors?
Measuring overhead is a hard thing to do. If you just want to measure
the total time in MPI, an analysis suite like MPE can probably give
you the information you desire. If you're looking to be really
accurate, there's also the overhead of copying data around to prepare
for sending / receiving (if your app does that sort of thing).
As for the difference in performance when using different numbers of
processors, many people use the metric that's most important: time to
completion. A couple of calls to MPI_Wtime will get you that
information. Of course, there's an argument about whether you should
scale the data size with the number of processors, but that's a
different story.
Hope this helps,
Brian
--
Brian Barrett
LAM/MPI Developer
Make today a LAM/MPI day!
|
