Thomas Dick, SciComp

Title:
Efficient Algorithmic Differentiation handling of linear algebra libraries

Abstract:

In many modern simulation codes efficient linear system solvers are a key factor for performance. These libraries use a variety of parallelization techniques to optimize performance, like MPI or different shared memory parallelizations. In this talk we will use the Spliss library developed by the DLR and used in CFD codes like TRACE as an example. We will discuss the challenges this context poses for Algorithmic Differentiation and explain a solution based on external functions.