Figure 6

Scaling properties of the different parts in phew obtained by restarting a cosmological dark matter simulation with \(\pmb{512^{3}}\) particles at redshift \(\pmb{z=0}\) . The top two panels show the runtimes of the different algorithmic blocks in phew. The peak patch segmentation and the saddle point search exhibit excellent scaling in the entire range of MPI tasks that we have tested. The merging in our test scales well up to ∼256 MPI tasks. The bottom panel shows the maximum number of sparse matrix elements over all MPI tasks compared to \(1/N_{\mathrm{tasks}}\) and rescaled to one at 32 MPI tasks. The increase seen in this number for of tasks is due to the growing load imbalance in terms of peaks per task and the increase in the surface to volume ratio of the domain segmentation. It explains the increase of the scaled runtime of the noise removal very well up to 512 tasks. The overall scaling of the algorithm is satisfactory up to 1,024 MPI tasks which is four times the number of CPUs the original simulation was run on.