Abstract of VIRIATO2 Project

This proposal aims at improving the scalability of the code Viriato, a unique framework that solves a reduced (4D, instead of the usual 5D) version of gyrokinetics applicable to strongly magnetized plasmas. Viriato is fully parallelized using domain decomposition with MPI; it is pseudo-spectral perpendicular to the magnetic field, and employs a high order upwind scheme for discretizing the equations along the magnetic field. A distinguishing feature of the code is its use of a spectral representation (Hermite polynomials) to handle the velocity space dependence. The Hermite representation converts a drift-kinetic equation for the electrons into a coupled set of fluid-like equations for each of the coefficients of the Hermite polynomials. In terms of numerical accuracy, this is rather advantageous: spectral representations are more powerful that grid-based ones, and it enables the deployment of the standard tools of CFD to deal with what would otherwise be a difficult kinetic equation. On the other hand, the scalability of the resulting algorithm, which uses extensively bi-dimensional Fourier transforms, becomes a non-trivial problem, whose improvement using MPI-3 shared memory concepts is the main goal of this proposal.