Looking at message traffic using Pablo
If you can coax an SDDF tracefile out of your system, then you can use
Pablo to get a statistical analysis of many different facets of your
program. At Cornell, we wrote a simple filter (mp2sddf) to convert
files written in IBM's native poe trace format into SDDF. There are
several ways to count up the MPI communications or user events in the
trace, but I like to use Pablo. You put together a bunch of Pablo
analysis widgets, with the tracefile going into the top of the graph:
![[A Pablo Configuration Graph]](Gifs/PabloConfig.gif)
Now, when you "run" this configuration graph, what comes out of the
bottom is the results of the analysis. For example, here I plotted
the total volume of communications as a function of time:
![[The resulting graph]](Gifs/PabloPlot.gif)
The plot shows a few communications at the beginning of the
run and a steady stream at the end of the run. You could also
display the volume of communications between individual nodes as an
led matrix, say.
(Back to main talk)