Collecting performance data on programs that call the MPI library is best done with the collect command. To collect data, insert the collect command and its arguments before the program name in the the mprun command:
mprun -np nproc collect collect-arguments program-name You can collect performance data on an MPI program using dbx, by inserting dbx collector commands in a script: mprun -np nproc dbx program-name < collection-script
You can collect performance data on running MPI programs by attaching dbx to each MPI process. However, the time taken to attach to the processes can affect the synchronization between the processes, and you cannot collect MPI tracing data unless you preload the Collector library before the MPI program starts running.
Data collection under MPI creates a separate experiment for each of the nproc MPI processes. The default experiment name is stem.n.er; n is the MPI rank of the process and stem is the stem of the experiment group name if one is given, otherwise it is test. If a name is explicitly given, the usual mechanism for naming experiments is used and there is no correlation between n and the MPI rank.
See also | |
---|---|
Experiment Names Experiment Groups MPI Tracing Data Collecting Performance Data on a Running Process |