Experiment to put system under many levels of stable load. Achieved by running 1,2,3...44 cores (4 threads each) with one of 7 kernels for 60s each.
Will run for approx. 5h total.
To run this experiment as designed, these additional components are required:
- Score-P installed
- Score-P ibm powernv plugin in
$LD_LIBRARY_PATH - Score-P metriq plugin in
$LD_LIBRARY_PATH module restore roco2-metricq-mlmust restore the used modules for the build; copy it from this directory:cp taurus_lmod_list ~/.lmod.d/roco2-metricq-ml
Recommended way to build, provided you are at the top-level directory of roco2:
module restore roco2-metricq-ml
mkdir build && cd build
SCOREP_WRAPPER_INSTRUMENTER_FLAGS='--user --openmp --thread=omp --nocompiler' SCOREP_WRAPPER=off cmake .. -DCMAKE_C_COMPILER=scorep-gcc -DCMAKE_CXX_COMPILER=scorep-g++ -DUSE_SCOREP=ON -DBUILD_TESTING=OFF -DENABLE_LOGGING=OFF -DP9_LONGRUN_METRICQ_SERVER='amqps://USER:PASS@HOST'
make SCOREP_WRAPPER_INSTRUMENTER_FLAGS='--user --openmp --thread=omp --nocompiler'
Note that you can use P9_LONGRUN_METRICQ_SERVER to specify a default metricq server.
Logging should be disabled, as the OMP barriers of the default logger intefered and stalled some processes up to 10s during trial runs.
Running make builds an accompanying slurm script, from the build directory it can be launched with:
sbatch -A ACCOUNT -- src/configurations/p9_longrun/run_slurm.sh
The result will be a directory beginning with scorep- and will be placed in your current directory.
The script will try to guess the correct metricq settings for the variables $SCOREP_METRIC_METRICQ_PLUGIN and $SCOREP_METRIC_METRICQ_PLUGIN_SERVER.
Just set them before invoking the script with sbatch, your settings will be kept.
Please note that SLURM sometimes has hiccups and requires additional parameters, especially --hint=multithread may be required manually.
Obviously the script is highly dependent on your local deployment, so adjustments may be required.