CitcomSX: Robust preconditioning in CitcomS via PETSc

The Citcom family of mantle convection codes are in wide spread use through out the geodynamics community. Since the inception of the original Cartesian version written in the early 1990's, many variants have been developed. Two important contributions were made by Shijie Zhong in the form of the parallel 3D Cartesian version and the parallel, full spherical version. Maintenance and support of CitcomS through CIG has seen a further increase in the development and usage of this particular version of Citcom. Such improvements have primarily been concerned with the introduction of new physics (rheology, compressibility), coupling with other software, including additional geologically relevant input/outputs and improved portability. Today, CitcomS is routinely used to solve mantle convection and subduction models with approximately one hundred million unknowns on large distributed memory clusters. Many advances have been made in both numerical linear algebra and the software encapsulating these concepts since the time of the original Citcom, however, the solver used by all Citcom software has remained largely unchanged from the original version. Incorporating modern techniques into CitcomS has the potential to greatly improve the flexibility and robustness of the method used to solve the underlying saddle point problem. Here we describe how PETSc (www.mcs.anl.gov/petsc), a flexible linear algebra package, has been integrated into CitcomS in a non-invasive fashion which i) preserves all the pre-existing functionality and ii) enables a rich infrastructure of preconditioned Krylov methods to be used to solve the discrete Stokes flow problem. The "extension" of the solver capabilities in CitcomS has prompted this version to be referred to as CitcomSX. We demonstrate the advantages of CitcomSX by comparing the convergence rate and solution time of the new Stokes solver with the original CitcomS approach. The BFBt preconditioner we utilise is robust and yields convergence rates largely independent of the element resolution and the viscosity contrast. Using this preconditioner, we can accommodate higher viscosity contrasts than were possible with the original CitcomS solver. The test problems used for this comparison include variable viscosity mantle convection (regional, full spherical) and slab subduction (regional).