1: \begin{abstract}

2:   We present a weighted \BFBT\ approximation (\wBFBT) to the inverse Schur

3:   complement of a Stokes system with highly heterogeneous viscosity.

4:   When used as part of a Schur complement-based

5:   Stokes preconditioner,

6:   we observe robust fast convergence %

7:   for Stokes problems with smooth but

8:   highly varying (up to 10 orders of magnitude) viscosities,

9:   optimal algorithmic scalability with respect to mesh refinement,

10:   and only a mild dependence on the polynomial order of high-order finite

11:   element discretizations (\discrQPdisc{k}{k-1}, order $k\ge2$).

12:   For certain difficult problems, we demonstrate numerically that

13:   \wBFBT\ significantly improves Stokes solver

14:   convergence over the widely used inverse viscosity-weighted pressure mass

15:   matrix approximation of the Schur complement.

16:   In addition, we derive theoretical eigenvalue bounds to prove spectral

17:   equivalence of \wBFBT.

18:   Using detailed numerical experiments, we discuss modifications to \wBFBT\ at

19:   Dirichlet boundaries that decrease the number of iterations.

20:   The overall algorithmic performance of the Stokes solver is governed

21:   by the efficacy of \wBFBT\ as a Schur complement approximation and, in

22:   addition, by our parallel hybrid spectral-geometric-algebraic multigrid (HMG)

23:   method, which we use to approximate the inverses of the viscous block and

24:   variable-coefficient pressure Poisson operators within \wBFBT.

25:   Building on the scalability of HMG, our Stokes solver achieves

26:   a parallel efficiency of 90\% while weak scaling

27:   over a more than 600-fold increase from 48 to all 30,000 cores of

28:   TACC's Lonestar~5 supercomputer.

29: \end{abstract}

30: