Abstract of MGTRI Project

In the general structure of a gyrokinetic code, a series of field solves appears at the end of each time step. Current methods are limited to topologically rectangular or polar grids. Several fluid codes were built last year around an ordered triangular grid, eliminating singularities at both magnetic axis and X-point. A parallel multigrid solver would greatly help turning these into world-standard massively parallel global codes, not least, able to resolve strongly nonlinear drift-wave turbulence and the crash of the tokamak edge-localised-mode (ELM) at those two points. Presently restricted to algebraic multigrid (AMG) solvers within the PETSc library, a home-grown geometric multigrid (MG) one would be better as it is not only more portable (no external libraries beyond MPI) but also more efficient since AMG is a more complicated method which is turned to when MG, which exploits coordinate cell isotropy, cannot be used. AMG is also harder to parallelise. We expect a successful MG solver for the elliptic field equations to parallelise to at least 256 cores for a case corresponding to ASDEX-Upgrade and, assuming capable weak scalability, to more than 1024 cores for an ITER case 4 times larger in linear dimensions (in both cases, we assume resolution of the collisionless plasma skin depth).