From: Nikita Andreev (lestat_at_kemsu.ru)
Date: Fri May 22 2009 - 06:03:30 PDT
> no problem. Sorry for not answering on your last mail yet. We are > currently in the testing phase for a new release which leaves very > little time to check on other thing. I will get back to you as soon as I > can. Actually I was referring to Dan Bonachea. :) > We don't compare different paths. In Cube you select the Metric > (Imbalance) first, then then a specific path, and then you can check the > distribution on the processes. I think I completely misunderstand this metric. I will take a look at Cube. > For MPI we are using a pre=instrumented library with wrapper functions. > As we don't measure the MPI implementation but the application we do not > look inside the MPI itself. After all there are quite a number of MPI > implementations out there that are not open source, so to which source > file do you want to map that, then? You don't need to dig in MPI functions itself but it would be helpful if you can say that for instance this MPI_Send-MPI_Recv pair has late sender issue and show in the code where these function were called (source file, line num). But if all MPI_* functions are from fake file "MPI" and without any begln/endln information you can't do that, can you? Regards, Nikita.