Our team has been developing a 5-dimensional gyrokinetic code (3 in position, 2 in velocity) for more than 20 years. It aims at simulating turbulence and transport in controlled fusion devices such as tokamaks. The code named GYSELA -- stemming from SEmi-LAgrangian GYrokinetic -- is written in Fortran 90 and uses an hybrid MPI/OpenMP parallelization approach.
It runs efficiently on several hundred thousand cores on standard CPU architectures: a relative efficiency of 85% on more than 524'288 cores has been recently achieved, and of 63% on 729'088 cores on the AMD EPYC architecture.
The largest production runs use over 200 billion grid points in phase space, and require several tens of millions of CPU hours.Yet, the code needs to be radically modified to adapt to the next generation of HPC architectures that target exascale (10^18 flops) and where GPU will be dominant.
To this aim, we have decided to rewrite it in modern C++, and to incorporate the state-of-the-art libraries and numerical tools that are currently developed in the framework of scientific collaborations, at the national and European levels. The new code, capitalizing on our 20 year experience, will be called GyselaX. You will be one of the main developer of this new code.
