logfile in parallel computing

[aleximos]

I have a question to the Logfile in parallel computations. I divided my mesh into 4 parts and checked the cpu time in the log file for specific load steps. There I noticed that step 100 in part 1 is already finished, when step 98  in part 2 is not.
I thought parts are exchanging information on ghost nodes to solve, so shouldnt part 1 wait for part 2 to finish?

[FEAP_Admin]

The problem is that the output is being buffered.  If you want you can force a flush just after the write to ilg in program/pmacr1.f .

[aleximos]

Thanks for the quick response!

So if I am not mistaken the log file only show the time needed for computation? Any waiting time without actually computing anything will be excluded, right?

As I have a script for performing the parallel run where there is also an output of the overall time needed for the calculation, there would be no more output needed inside of feap.

Do you think this information about different time needed for partitions could be used to partition the model in an optimized way? For example one runs a calculation, checks where the most time is needed and changes some of the elements in this partition to another one with less time needed for calculation? Or is there any other factor influencing the computational time which I need to keep in mind here?

[FEAP_Admin]

The best way to look for balancing errors is to use -log_summary in PETSc.  This will give an accurate picture of any unbalanced loads.  And provide you information on its causes.