I'm timing a Fortran/MPI code, and I got the following timing
information for one of 4 processes (the other 3 processes gave
similar results):

wallclock time  26.2 sec.
cpu time        23.4 sec.
MPI_Barrier      7.8 sec.

I had assumed that MPI_Barrier would accumulate wallclock time but not
cpu time, but this assumption is not consistent with the above
timings, since the wallclock and cpu times differ by less than the
time spent in MPI_Barrier.  I ran this case on an IBM p690 under AIX.
Does MPI_Barrier accumulate cpu time?  Is this operating-system
dependent?  If so, does anyone know whether it's true for AIX, linux,
and IRIX?  Thanks for any help you can offer.

Doug Sondak

In article <br5svs$pk0$1@news3.bu.edu>, Douglas Sondak <sondak@bu.edu> wrote:

>I had assumed that MPI_Barrier would accumulate wallclock time but not
>cpu time,

That's implementation-dependent. On many low-latency implementations,
polling is used. If you don't have any other useful work to do, that's
a good idea. Sometimes it's switchable via an environment variable or
flag.

All of the platforms you asked about have multiple MPI
implementations.  I would suggest reading the docs and observing
reality to find out what they do.

-- greg