Page Comparison

...

Task and Submission

1. Use the run input in the POT3D package’s under testsuite/isc2023 folder.

2. Run POT3D with isc2023 input on both PSC bridges-2 and FAU Fritz CPU clusters using 4 nodes.
   Experiment with number of ranks per socket/numa domains to get the best results.
   Your job should converge at 25112 steps and print outputs like below:

   Code Block
   ### The CG solver has converged. Iteration: 25112 Residual: 9.972489313728662E-13

3. Profile a run.

   1. Use any of the remote clusters to run an MPI profile (such as IPM profile or any other profiler) for a run using 4 nodes with full PPN.

   2. Submit the profile as a PDF to the team's folder.

   3. In your presentation, also indicate the 3 main MPI calls that are being used and their run times, as well as the total MPI time for the test.

4. Bonus task: Run POT3D on the PSC cluster using the V100 GPU partition.

   1. Use only 4 GPUs for the run. It is recommended to use one rank per GPU.

   2. Submit the results to the team's folder.

   3. NOTE: To compile and run POT3D with the nvfortran compiler, you must load and/or build the HDF5 library compiled with nvfortran. The code is known to work with HDF5 1.8.21 (http://portal.hdfgroup.org/display/support/HDF5+1.8.21 )

   4. An example build script for POT3D with the NVIDIA compiler can be found in build_examples/build_gpu_nv22.3_ubuntu20.04.sh

   5. Note that you do NOT need to enable the cusparse option because the test case is not set up to use the algorithm that requires cusparse. Therefore, if linking cusparse is causing difficulties, you can change the build script line POT3D_CUSPARSE=1to POT3D_CUSPARSE=0.

5. Submission and Presentation:
   - Submit all your build scripts, run scripts, inputs, and output text files (pot3d.dat, pot3d.out, timing.out, etc.)
   - Do not submit the output HDF5 data or the source code.
   - Prepare slides for the team’s interview based on your work for this application.

...