A Pre-Processing Interface for Steering Exascale Simulations by Intermediate Result Analysis through In-Situ Post-Processing

Abstract

Today's simulations are typically not a single application but cover an entire tool chain. There is a tool for initial data creation, for partitioning this data, for actual solving, for afterward analysis, for visualisation. Each of these tools is separated and so is the data flow. This approach gets unfeasible for an exascale environment. The penalty for every data movement is extreme. A load balance optimised for the solver most likely does not hold true for the complete run time. Compared to tweaking each part, our solution is more disruptive: Merge the tools used and thereby improve overall simulation performance. In this paper, we provide a first step. We combine pre-processing, simulation core and post-processing. We outline our idea of a general pre-processing interface steering the overall simulation and demonstrate its applicability with the sparse geometry lattice-boltzmann code HemeLB, intended for hemodynamic simulations. The tasks of the interface start with partitioning and distributing simulation data. Main aspect is to balance load and communication costs of the entire simulation by using the costs of each simulation part. Measurement of these costs for each simulation cycle makes possible further performance improvement: Data can be redistributed in between cycles in order to achieve a better load balance according to these costs. Additionally, our interface offers the possibility for other easy-to-integrate extensions covering, e.g., an automated mesh refinement or fault tolerance awareness. Finally, we investigate the applicability of our interface within HemeLB. We exploit the integrated partitioning methods and especially consider latest HemeLB-specific in-situ result analysis and visualisation methods.