ShellSetHPC – Parallel dynamic neotectonic modelling

Geodynamic forward models are typically computationally expensive, they generally draw on a large set of input and testing data and will normally have quite large model domains in order to ensure that the results are practical and applicable. To reduce simulation time many programs will use some form of parallel computing in their calculations, e.g., programs may use the message passing interface (MPI) communication protocol or OpenMP API to partition the domain into reasonable chunks divided between multiple processes, this can improve performance by having multiple processes solve smaller parts of a large problem in parallel.

Geodynamic inverse modelling involves using known phenomena to constrain a model, or set of models, to improve knowledge of unknowable or unmeasurable parameters involved within a dynamic system. For example, one could use measured GPS velocities and seafloor spreading rates to infer changes to mantle convection or combine GPS velocities with near-surface temperatures to monitor the growth of magma chambers.

Since geodynamic inversion involves running multiple constrained forward models it naturally suffers from the performance issues linked to the forward models themselves, briefly: a poorly performing forward model will mean a poorly performing inverse model. More than that, the inverse model must perform multiple (the more the better) forward models with updated parameters. These requirements, a good enough forward model, and an efficient method of performing multiple models in parallel while optimizing the desired parameters, leads to an obvious conclusion: to perform forward models in parallel while searching for an optimal model.

To this end we present a geodynamic inversion model, which we call ShellSetHPC, which uses a combination of existing, well known, and robust software to model the neotectonics of planetary lithosphere. This is further combined with an efficient, de-centralised random search algorithm able to generate testable models within a user defined N-dimensional space. This algorithm is also able to launch multiple models in parallel thanks to an MPI framework. The forward model makes use of Intel’s thread safe math kernel library (MKL) to solve the linear system in parallel. This hybrid approach lends itself to use on high performance computing (HPC) machines which would allow a more complete utilisation of these features.

In this work we will present scaling results on an HPC cluster and compare these with results obtained from more typical search algorithms. All tests are performed within a realistic geological setting, the results of which will be used to gather insight into the performance of the driving model generators and their settings.