July 11 – Jose Fonseca: Towards scalable parallel adaptive simulations with ParFlow
Jose Fonseca: Thursday 11 July at 11:00, A415 Inria Paris. The accurate simulation of variably saturated flow in a porous media is a valuable component in understanding physical processes occurring in many water resources problems. Such simulations require expensive and extensive computations and efficient usage of the latest high performance parallel computing systems becomes a necessity. The simulation software ParFlow has been shown to have excellent solver scalability for up to 16k processes. In order to scale the code to the full size of current petascale systems, we have reorganized its mesh subsystem to use state of the art mesh refinement and partition algorithms provided by the parallel software library p4est. Evaluating the scalability and performance of our modified version of ParFlow, we demonstrate weak and strong scaling to over 458k processes of the Juqueen supercomputer at the Jülich Supercomputing Centre. In the first part of the talk we will briefly present the algorithmic approach employed to couple both libraries. The enhanced scalability results of ParFlow’s modified version were obtained for uniform meshes. Hence, without explicitly exploiting the adaptive mesh refinement (AMR) capabilities of p4est. We will finish this first part presenting our current efforts to enable the usage of locally refined meshes in ParFlow. In an AMR framework. In such case, the finite difference (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 talk presenting a block preconditioner for saddle point problems arising from a MFE that retains its robustness in the case of locally refined meshes.