Performance of Massively Parallel Multigrid on Hybrid Tetrahedral Grids

We discuss the performance analysis of massively parallel matrix-free geometric multigrid solvers on hybrid tetrahedral grids (HyTeG). Beyond h-independent algebraic convergence, the notion of textbook multigrid efficiency (TME) suggests a concrete bound for the computational complexity of the iterative solver. In this talk we consider a performance model based on TME applied to a monolithic multigrid algorithm targeting the Stokes problem. The performance model is extended to a massively parallel setting, solving saddle point systems with more than a trillion (> 10^{12}) unknowns on roughly 150,000 processes.