6.1 Profiling

The timing of a serial (single processor, single threaded) program is usually measured as the ``real CPU time'' (the time which is spent executing commands of the program) by the kernel of the operating system. Thus, idle times during which the program is ``sleeping'' and waiting for other processes or i/o operations to complete are not taken into account.

On the contrary, the timing of a parallel program has to be done in terms of the elapsed time between its invocation and termination (``wall clock time''). Idle times of one of the different instances of the program indicate a load balancing problem, which may lead to a less than optimal scaling. As a result, careful profiling requires, that all instances have an equal environment on their processor (i.e. no other processes producing a high load should be running) and all timings have to made in terms of ``elapsed wall clock time''.

Accurate profiling of the application is supported by several PETSc features, which are activated with command line arguments for the PETSc application:

• -log_summary prints a summary of timing data and performance statistics at the end of the program. The impact on the overall performance of the program is very small.
• -log_info prints detailed information about data structures, solver performance and i/o operations. Printing these data on stdout or to a file slows down the program considerably, but it provides accurate information about specific parts of the program. If this option is selected, the execution of all subroutines of the micromagnetics application is timed individually.