6.7 Summary and outlook

We presented several application scenarios to show the applicability and benefits of dynamically adaptive grids. Based on an analytic benchmark, we showed the correct implementation and that over 96% cells can be saved with a dynamically adaptive grid. Then a Tsunami simulation was implemented to show the applicability of the dynamic adaptivity within a realistic scenario. With a relatively small benchmark scenario, the computation time for the simulation was already improved by a factor of 6.6 with results of higher accuracy compared to the simulation on a regular grid. Assuming, that the results with the dynamically adaptive simulation are sufficiently accurate, the runtime improvement is larger than a factor of 54. For the data analysis with an offline processing, we tested several output backends for online and offline processing. Here, online processing allows direct visualization of the simulation as well as interactive steering methods. We further developed several ways of writing output data to persistent memory. Here, writer tasks result in only 4% loss in performance compared to a simulation which is not writing any output data. Finally, we presented our extensions for simulations on the sphere and a multi-layer discretization.

There are a couple of issues that deserve further investigation in the future. Among others, these are given as follows:

• Cluster-based local-time stepping:
  With the naming “clustering” originating from the cluster-based local-time stepping idea, this is obviously one possible utilization of our approach. Since our cluster-based approach is similar to block-structured grids (see Sec. 5.10.5), we expect that an extension to cluster-based local-time stepping can be accomplished in a similar way as in the PeanoClaw [UWKA13] framework.
• Resiliancing:
  The independency of clusters allows an efficient duplication and forwarding of one or more clusters to other compute nodes for resiliancing. A replacement of a computation node can then be accomplished by initializing the simulation at another node, based on the duplicated clusters and a reconstruction of RLE adjacency information only on the cluster-adjacent MPI ranks.
• Dynamically changing simulation data:
  Considering our GUI which also offers interactive steering methods such as setting parameters during the simulation’s runtime and modifying the cell data and grid structure, this yields further applications. Some of them are e.g. the interactive testing of flooding scenarios or the simulation of dynamic earthquake-induced displacement data.