2020-07-29 –, Green Track
We present an iterative and massively scalable 3-D multi-GPU inversion workflow using Julia for coupled multi-physics processes in Earth Sciences. We introduce an adjoint framework for the two-phase flow equations, assess the unknown porosity field reconstruction in 3-D and discuss the performance evaluation.
The adjoint-based multi-physics inversion framework we present enables the development of efficient and massively scalable 3-D multi-GPU solvers with application to optimisation problems.
We use an iterative matrix-free pseudo-transient approach and the finite difference method to solve the forward and the adjoint coupled two-phase flow equations. We achieve efficient calculations of the pointwise gradients of the flow solution with respect to the porosity. We then use the gradients in a gradient descent method to reconstruct the pointwise porosity in 3-D.
We assess the performance of the 3-D memory-bounded solvers using a simple effective memory throughput metric. We finally discuss how the overlap of computations with MPI communications permits us to achieve a close to optimal parallel efficiency. We rely on the ParallelStencil and ImplicitGobalGrid packages for high-performance stencil-based calculations and optimal distributed memory parallelisation.
Co-authors - Georg Reuber, Samuel Omlin
Why to perform serial tasks while nature is mostly parallel? I am geoscientist with strong interest in high-performance computing (HPC), two-phase flow and ice dynamics. GPUs and supercomputers permit to investigate previously unsolvable problems at unprecedented resolutions.
Computational Scientist and responsible for Julia Computing at the Swiss National Supercomputing Centre (CSCS), ETH Zurich