Advanced HPC methods for large-scale sensitivity analysis

Nonlinear numerical control is rightfully considered to be one of the most difficult engineering problems to tackle, in terms of both practical implementation and time-to-solution. It requires time-stepping numerical models for simulating the trajectory of the system, adjoint models for sensitivity analysis and matrix-free iterative solvers to produce the solution field of the inverse problem. The field of high-performance computing (HPC) provides computational tools and practices which enable the deployment of numerical applications on computational clusters, supercomputers and cloud computing facilities. This article presents a set of practical methods aimed at accelerating and parallelizing computation for sensitivity analysis in a large-scale 4D-Var data assimilation setting.