On 5/27/09 12:42 AM, Supalov, Alexander wrote:

Hi everybody,

I added my comments to the proposal for lack of feedback per email. Please consider at F2F. Slight correction to item 5:

Reads "Likewise, the MPI-stall can probably be fairly well statistically approximated by mapping PC to the progress engine routines and its dependants."

Should read "Likewise, the MPI-stall can probably be fairly well statistically approximated by mapping PC to the progress engine routines and its dependants, except memcpy and equivalents, and wire access routines".

Yes, that's true, but it implies that the tools would know a lot of detail about the MPI implementation. We wanted to avoid that, since every MPI implementation (and the Sun profiling tools support seven different versions) is different. In the Sun MPI profiler we also (by default, although it can be changed) suppress all call frames below the first user call into MPI, to avoid overwhelming the user with details that he or she can do nothing about. Doing so removed more than half the functions from the function list data for a run. Marty