omp-critical experiments

The omp-seq-reduction and omp-par-reduction codes were run on a single 48-core node of MareNostrum-4 at Barcelona Supercomputing Center for a range of thread counts from 1 to 48 and the results compared to those of the original omp-master code. The problem size in all cases was specified by Ni=500, Nj=500, Nk=500 and Nt = 20. The run times and the average IPC (Instructions Per Cycle) values for all three versions are shown here.

Clearly, the new versions of the code show much better performance and have much better avaerage IPC values, with those for the omp-seq-reduction version being marginally better. In order to compare the run times of the omp-req-reduction and omp-par-reduction kernels, these times are plotted below without the much slower master kernel. The omp-seq-reduction version can now be seen to be slightly faster at the lower thread counts. This may of course vary with details of the simulation.