K-Means Benchmark Results

Each system has two subdirs: OLD and NEW

OLD

Legacy runs performed on K-MEANS code provided by RWTH (un-modified code runs or with minor changes such as MAQAO probes and added a second parallel for clause before the centroid position update, although not measured by the probes, only for slightly better performance)

NEW

Modified K-Means code with probes, measuring both point-to-centroid assignment and centroid position update. Now both loops are scheduled statically by OMP (default is 'auto' i.e. implementation defined, see OMP 6.0 doc section 13.6.3). Tested versions are O2, O3 and Ofast/ffast-math with more precise arch-tuning flags (march,mtune/mcpu,xHost). OMP_WAIT_POLICY env. var. is set to ACTIVE in order to force G++ to change from its PASSIVE default and for MAQAO to be able to sample.