High Performance Computing

Parallel Profiler: Jumpshot

A profiler for parallel code will inform the developer with more information on the message passing artefacts of ones source. An example serves as the best way to illustrate this. Copy /opt/score5.8.1/score-src/runtime/mpi/mpich-1.2.0/src/examples/basic/fpi.f to somewhere convenient. The issue

mpif77 -mpilog fpi.f
mpisub 3x2 a.out
This will generate a file called fpilog.clog which we need to convert to SLOG format:
clog2slog fpilog.clog
Start the profiler:
logviewer fpilog.slog

An extra X window will now appear with a graph in it

image of Jumpshot running

From this you can easily see how many calls each MPI call has made and the time to do so. Clicking on DISPLAY will open a second windows where the developer can see more timing statistics by clicking on the bars themselves or clicking on the legend towards the bottom.