Scalable parallel simulation of variably saturated flow

Abstract

In this thesis we develop highly accurate simulation tools for variably saturated flow through porous media able to take advantage of the latest supercomputing resources. Hence, we aim for parallel scalability to very large compute resources of over 10⁵ CPU cores.<br /> Our starting point is the parallel subsurface flow simulator ParFlow. This library is of widespread use in the hydrology community and known to have excellent parallel scalability up to 16k processes. We first investigate the numerical tools this library implements in order to perform the simulations it was designed for. ParFlow solves the governing equation for subsurface flow with a cell centered finite difference (FD) method. The code targets high performance computing (HPC) systems by means of distributed memory parallelism. We propose to reorganize ParFlow's mesh subsystem by using fast partitioning algorithms provided by the parallel adaptive mesh refinement (AMR) library p4est. We realize this in a minimally invasive manner by modifying selected parts of the code to reinterpret the existing mesh data structures. Furthermore, we evaluate the scaling performance of the modified version of ParFlow, demonstrating excellent weak and strong scaling up to 458k cores of the Juqueen supercomputer at the Jülich Supercomputing Centre.<br /> The above mentioned results were obtained for uniform meshes and hence without explicitly exploiting the AMR capabilities of the p4est library. A natural extension of our work is to activate such functionality and make ParFlow a true AMR application. Enabling ParFlow to use AMR is challenging for several reasons: It may be based on assumptions on the parallel partition that cannot be maintained with AMR, it may use mesh-related metadata that is replicated on all CPUs, and it may assume uniform meshes in the construction of mathematical operators. Additionally, the use of locally refined meshes will certainly change the spectral properties of these operators.<br /> In this work, we develop an algorithmic approach to activate the usage of locally refined grids in ParFlow. AMR allows meshes where elements of different size neighbor each other. In this case, ParFlow may incur erroneous results when it attempts to communicate data between inter-element boundaries. We propose and discuss two solutions to this issue operating at two different levels: The first manipulates the indices of the degrees of freedom, While the second operates directly on the degrees of freedom. Both approaches aim to introduce minimal changes to the original ParFlow code.<br /> In an AMR framework, the FD method taken by ParFlow will require modifications to correctly deal with different size elements. Mixed finite elements (MFE) are on the other hand better suited for the usage of AMR. It is known that the cell centered FD method used in ParFlow might be reinterpreted as a MFE discretization using Raviart-Thomas elements of lower order. We conclude this thesis presenting a block preconditioner for saddle point problems arising from a MFE on locally refined meshes. We evaluate its robustness with respect to various classes of coefficients for uniform and locally refined meshes.