I have recently tried to estimate the parallel efficiency of ADDA, but it was
not that easy. In principle, timing section of the log file contains almost all
required information, like communication time for different parts of the run.
However:
1) Total communication time is not available. In particular, when more than 1
solution for the internal fields is needed (e.g. 2 incident polarizations),
communication time is indicated only for the last solution. The only workaround
is to scale this times by ratio of the total time of for the internal fields to
that for the last solution.
2) Anyway, one need to do some calculations to get the numbers. So it would be
convenient if simple numbers (like "83%") would be given explicitly in the end
of the timing section (say, overall and for the internal fields).
Original issue reported on code.google.com by
yurkinon 2 Sep 2013 at 3:29