OpenACC/MPI Demos

These demos use OpenACC and MPI to run SAXPY in various contexts.

Contents

• saxpy_acc runs a SAXPY in parallel on a single GPU.

• saxpy_mpi runs SAXPY on CPU with multiple MPI processes. The root allocates an array for the global problem, then scatters local subarrays to the other ranks. Then, each rank computes SAXPY on its subarray. Finally, the subarrays are gathered back onto the root to check the results. The length of the global problem must be evenly divisible by the number of MPI ranks.

• saxpy_acc_mpi runs SAXPY with multiple MPI processes. Each process offloads its local work to GPU using OpenACC kernels. The MPI scheme is the same as saxpy_mpi, and the length of the global problem must be evenly divisible by the number of MPI ranks. The number of NVIDIA-capable GPUs can be arbitrarily >= 1; each rank's kernels will be round-robin scheduled on the available GPUs.

Building

Ensure that the mpif90 in your path uses pgfortran. Then simply run make all.

Running

MPI programs are run as usual, e.g: mpirun -np 2 ./saxpy_acc_mpi

Profiling

The nvprof users' manual describes how to profile and visualize MPI sessions:

• MPI Profiling With nvprof
• Import Multi-Process nvprof Session

For example, you could run nvprof as the following, where %p is substituted with the corresponding process ID:

mpirun -np 2 nvprof --export-profile saxpy.%p.prof ./sgemm_acc_mpi

After copying the *.nvprof files to your local computer, you can import and visual them in nvvp.