Abstract of RAGOUT Project
Preparing an application for a transition from petascale to pre-exascale systems requires large investments in terms of software research and development. Targeting billions of computing elements is disruptive and will require new core schemes and algorithms to achieve extreme parallelism. In particular the shift to a different programming model, relying on a runtime system, is necessary. To this end, the refactoring and rewriting of the GYSELA application is scheduled. GYSELA is a 5D gyrokinetic code that solves the coupled set of Vlasov and quasineutrality equations for the entire distribution function (full-f) in global geometry [V. Grandgirard et al., Comp. Phys. Com. (2016) 35, DOI: 10.1016/j.cpc.2016.05.007]. The numerical scheme is semi-Lagrangian, backward in time. With this code, we achieved a relative efficiency of 91% on the full Juqueen machine composed of 458,752 cores [http://www.fz-juelich.de/ias/jsc/EN/Expertise/High-Q-Club/_node.html].
However, as HPC is fueling breakthroughs in science, redesigning a fresh gyrokinetic code shall provide a tool that goes beyond topical modeling capabilities. To this end, we intend to assemble novel components to achieve a good representation of the plasma from the core up to the scrape-off layer, which is quite new. In this respect, we plan to use a multi-resolution approach in space involving heterogeneous mesh refinement and requiring dynamic load balancing techniques. The goal of this project is to provide a strong code basis to build up a new semi Lagrangian code. The main components provided by this project will be: (1) efficient 3D and 5D data structures with support for multi-resolution (we will investigate a few mesh refinement schemes), (2) efficient MPI communication schemes adapted to these data structures and (3) task programming to handle computational heterogeneity and overlap communication and computations. The design of the main algorithms and numerical schemes of this code will not be covered by this project, other people will do it during the course of the present project.