OpenMP Tasking and MPI in a Lattice QCD Benchmark

Beginning with an existing well-optimized lattice quantum chromodynamics solver using OpenMP+MPI, we develop two task-based implementations, one with OpenMP tasking and one with hand-coded “untasking”. We achieve better overlap of MPI communication and computation with both methods, and expose some performance issues in OpenMP tasking. Both task-based implementations outperform the original implementation when strong scaling.