Make sure the repo is cloned with the --recursive flag, e.g.

git clone --recursive https://github.com/kabicm/lu

To build and run, do the following:

mkdir build && cd build
CC=gcc-9 CXX=g++-9 cmake -DSCALAPACK_BACKEND=<scalapack-library> .. # or whatever version of gcc compiler you have
# <scalapack-library can be: MKL, CRAY_LIBSCI or CUSTOM>
make -j
mpiexec -np <num MPI ranks> ./lu -N <global matrix size> -b <block size> --p_grid=<prow>,<pcol> -r <num of repetitions>

# Example for LU:
mpirun -n 4 ./lu -N 1200 -b 128 --p_grid=2,2 -r 2
Warning: using only 4 out of 5 processes.
==========================
    PROBLEM PARAMETERS:
==========================
Matrix size: 1200
Block size: 128
Processor grid: 2 x 2
Number of repetitions: 2
--------------------------
TIMINGS [ms] = 353 186
==========================

# Example for Cholesky (output is structured in the same way):
mpirun -n 4 ./cholesky -N 1200 -b 128 --p_grid=2,2 -r 2

Enter the params you want to work with into scripts/params.ini. Now, move to the source folder and generate the .sh files by running python3 scripts/generate_launch_files.py. If you only want to generate scripts for lu, you can pass the argument algo lu. For only cholesky on the other hand, pass algo chol. If no argument is given, both are generated. You can specify the output folder for the benchmarks with --dir <path_to_folder>. It will default to ./benchmarks.

After having generated the files, run python3 scripts/launch_on_daint.py. It will generate allocate nodes for each processor size, at the moment using the heuristic that we have n nodes with 2n ranks. If you launch very large jobs, perhaps you have to change the runtime in python3 scripts/generate_launch_files.py or in the bash scripts directly (it defaults to 2 hours at the moment). Each .sh file contains all jobs for one specific processor size.

• Marko Kabic ([email protected])
• Tal Ben Nun ([email protected])
• Jens Eirik Saethre ([email protected])
• Andre Gaillard ([email protected])