Exercise: MiniMD Trace Collection

Build the MiniMD without likwid:

$ cd ~/Tools-Material/MINIMD
$ make

Run the uninstrumented MiniMD proxy app with 4 ranks and 10 threads each (ensure, that you are in the 'MINIMD/data' directory):

$ cd data
$ OMP_NUM_THREADS=10 mpirun -n 4 ../miniMD-ICC -t 10 --half_neigh 1

Build with modified Makefile configuration:

$ cd ..
$ make TAG=ICC.scorep
$ cd data

Run instrumented MiniMD proxy app:

$ export SCOREP_EXPERIMENT_DIRECTORY=scorep-minimd-4x10-profile
$ OMP_NUM_THREADS=10 mpirun -n 4 ../miniMD-ICC.scorep -t 10 --half_neigh 1

Examine scoring:

$ scorep-score scorep-minimd-4x10-profile/profile.cubex 

List also individual functions causing trace buffer requirements:

$ scorep-score -r scorep-minimd-4x10-profile/profile.cubex 

Generate an initial filter file base on default heuristic:

$ scorep-score -g scorep-minimd-4x10-profile/profile.cubex

Examine filter file:

$ cat initial_scorep.filter

Apply filter file to determine expected reduction:

$ scorep-score -f initial_scorep.filter scorep-minimd-4x10-profile/profile.cubex

Apply filter file to measurement and re-run MiniMD:

$ export SCOREP_EXPERIMENT_DIRECTORY=scorep-minimd-4x10-filtered
$ export SCOREP_FILTERING_FILE=initial_scorep.filter
$ OMP_NUM_THREADS=10 mpirun -n 4 ../miniMD-ICC.scorep -t 10 --half_neigh 1

Re-examine effect of filter file:

$ scorep-score scorep-minimd-2x10-filtered/profile.cubex

Enable trace file collection:

$ export SCOREP_EXPERIMENT_DIRECTORY=scorep-minimd-4x10-tracing
$ export SCOREP_ENABLE_TRACING=true
$ export SCOQREP_TOTAL_MEMORY=23MB
$ OMP_NUM_THREADS=10 mpirun -n 4 ../miniMD-ICC.scorep -t 10 --half_neigh 1

Examine experiment result:

$ ls -la scorep-minimd-4x10-tracing

"Switch to XFCE Dekstop" and start Vampir:

$ vampir scorep-minimd-4x10-tracing/traces.otf2

Questions:

• What functions are filtered out of the measurement? Does this list make sense in a parallel performance engineering context?
• How much overhead was there in the initial measurement? Is this an acceptable value?
• How much overhead was there in the filtered measurement? Is this an acceptable value?