Each of the four implementations, as well as the existing MPI_Alltoallv and point-to-point implementation were run on the HECToR Phase 2a and Phase 2b systems for each of the three benchmark systems, CNT40, Silicon and CNT80. The results are shown below. In all graphs, the X axis shows the number of cores, and the Y axis the speedup relative to the original MPI_Alltoallv code on 1 core.