Open Issues

While we think that the proposal outlined so far is reasonably complete, there are additional issues which need further discussion. Here, we try to briefly summarize the issues we are aware of:

• Do we need instrumentation for the OpenMP ORDERED and FLUSH?

• Do we need instrumentation of single iterations of PARALLEL DO (parallel for) loops? This would potentially allow the influence of OpenMP loop iteration scheduling policies to be measured, but only when iterations are of a sufficient granularity.

• Do we need a way to allow users to pass additional information to the POMP interface? Besides using POMP conditional compilation (see Section 3.6), we might consider extending the !$OMP INST BEGIN and END directive (see Section 3.4) to optionally allow to pass the address of a user variable:

  

  The pomp_begin and pomp_end routines would have an additional void* typed second parameter which would be NULL if the user did not specify a variable. This way, the user could pass additional information to the POMP API functions, which would be ignored in general implementations but could be utilized in (user-supplied) special versions. It is also possible to allow passing user information to the other (all) POMP API functions, but this requires further directives / pragmas.

• Do we need additional OpenMP runtime level instrumentation or is it enough to observe OpenMP behavior on the source code level?