Debugging and Timing MPI Programs
- Debugging
- Timing
- SGI Origin 2000
The TotalView Multiprocess Debugger is available. See the
Performance
Tools page for general information on TotalView, and follow the link
here for specific information on debugging MPI programs.
- Within the code
To get an idea of how much CPU time was used in various parts of your code so
that you can determine where you need to tune things, you can use
dtime or etime for Fortran codes. However, the timings
will be per thread when you do this so you'll need to either print the timings
from each thread or use an MPI_gather() operation to collect the times from
all of the nodes. If you are programming with C you can use the times()
system call which may be easier to use for C programmers.
In C, the routine MPI_wtime() gives
you the wallclock time used by your MPI job. This gets you the amount
of wallclock time that your job requires which will allow you to
calculate speedup. You should be warned that speedup measurements are
only accurate in the dedicated queues and you should not take too much
stock in the effective speedup you see in the batch queues due to the
range of load conditions and context switching that occurs there.
- Outside the code
- SGI Origin
/bin/time or timex do not work with SGI native
MPI.
As a workaround, a close approximation can be obtained by using
the perfex command as follows:
mpirun -np n perfex -y -o filename a.out
Note that the time obtained from perfex is the sum of all the MPI
processes.
