[mpich-discuss] Logging MPI Collectives

[Pavan Balaji]

On 02/15/2010 08:51 PM, sanath kumar wrote:
> Oh I understand. I read through the logging.pdf file. RLOG is
> independent of MPE right? My doubt is if RLOG can do the job of MPE,
> what is the need for MPE? Can you point out some differences between the
> both? Thank you.

Yes, rlog and MPE are independent. rlog is for internal profiling of
MPICH2 (inside the library), while MPE profiles at the MPI call level.
So, both have their own specific purposes. The common aspect is that the
logs generated by both these profilers can be converted to slog2 format
which is viewable by Jumpshot.

There are some differences in their implementations as well. For
example, rlog does in-memory profiling. Which means that it can't
profile very long running applications. There are some other differences
in the details as well (such as rlog doesn't have any inter-process
synchronization, so you get independent log files, which might not be
time synchronized). You need to view these as two independent tools.

If you are just looking to extend MPE to add additional profiling,
instead of using rlog, you'll need to add more states (see the calls to
MPID_MPI_STATE_DECL and MPID_MPI_FUNC_ENTER in MPIR_Bcast (bcast.c:981)
for example). But my suggestion is still to just use rlog instead.

 -- Pavan

-- 
Pavan Balaji
http://www.mcs.anl.gov/~balaji