Parallel programs are more difficult to understand than sequential ones. It is common to conceive of a parallel algorithm, implement it, and then be puzzled by its disappointing performance, even though the program executes correctly. What is then needed is an instrumentation of the program, so that data can be collected that will lead to an understanding of the program's behavior.
The difficulty is that instrumenting all of the system calls is likely to be cumbersome and error prone, while any specific profiling mechanism provided by the vendor of the parallel computer or the parallel library is unlikely to be just exactly what a user wants. This is particularly the case if the user needs to tailor the profiling data to a specific application, such as when custom run time or post mortem graphical displays are used.