Computational challenges in Peta-scale fusion plasma turbulence simulations

Idomura, Yasuhiro   

Fusion plasma turbulence simulations based on a five dimensional (5D) gyrokinetic model is one of the most demanding fields in computational fusion science. This paper presents novel parallelization techniques such as multi-dimensional domain decomposition on hierarchy of hybrid MPI/OpenMP networks and communication overlap techniques using a heterogeneous OpenMP model, which are developed for a gyrokinetic toroidal 5D Eulerian code GT5D. Performance tests on Peta-scale platforms demonstrate that the parallelization techniques enable us to access Peta-scale computing beyond 100 k cores, and open a new physics regime of fusion plasma turbulence simulations. Taking advantage of high parallel efficiency and sustained performance of GT5D, the plasma size scaling of turbulent energy transport is first addressed using full-f simulations with experimentally relevant conditions.