Many parallel algorithms become slower if you make them use too many cores, as they spend most of the time in communication, and very little in actual computations. Maybe the parallel systems would have been faster on a smaller number of cores than 128?