README-multigpu.md

Heat equation with multiple GPUs

Your task is now to combine the MPI parallelization as described for the CPU-only code with the OpenMP offloading.

You can base your work on the hybrid MPI + OpenMP code and the previous work on offloading with a single GPU.

In order to achieve a working multi-GPU code, you should:

1. Assign MPI tasks to devices
2. Alternatively, either
   • a) Copy the data between host and device before and after the MPI communication, or
   • b) Pass device pointer to MPI routines
3. Use OpenMP offload constructs in the evolve routine

Note: If opt for 2b then

• Set environment variable export MPICH_GPU_SUPPORT_ENABLED=1 prior to compiling abd running.
• Run the program with srun -n <ntasks> --gpus <ntasks> <rest of the flags> ./heat_offload.