Abstract of ERO2 Project
ERO2.0 is an established code for modelling plasma-wall interaction and global material migration in fusion devices [1]. The migration is simulated by following the 3D trajectories of Monte-Carlo test particles, which are representative of the eroded material, in the plasma edge. Acceleration by electromagnetic fields is taken into account, as well as atomic and molecular processes (ionisation, recombination, dissociation) and collisions (drift, diffusion). To obtain accurate trajectories in the sheath and impact velocities of the test particles, the 3D gyro-orbits are resolved (instead of applying the guiding-centre approximation). Since large numbers (106) of test particles are required to reduce Monte-Carlo noise, ERO2.0 is parallelised using a hybrid MPI/OpenMP scheme and routinely executed on the JURECA supercomputer [2]. Due to the trace impurity approximation underlying the model, trajectories can be calculated independently of each other in parallel by individual OpenMP threads. Since trajectory lengths (proportional to the associated work load) can vary by orders of magnitude, a dynamic load-balancing algorithm is used. This minimises the idle time of CPU cores, but requires extra communication between them. To make the simulations efficient in terms of computing resource usage, fine-tuning is required that depends on the application case as well as on the supercomputer architecture. This involves a proper choice of 1) the number of OpenMP threads used per MPI task, and 2) the size and frequency of MPI messages exchanged during the load balancing. For this project, we propose to use HLST resources for porting and tuning ERO2.0 on Marconi-Fusion. In particular, the above-mentioned parameters should be optimised for one well-known reference application case. An efficient parallel scaling should be demonstrated, preferably reproducing or excelling the one measured for JURECA.
[1] J. Romazanov et al., 2017 Phys. Scr. 2017 014018
[2] Jülich Supercomputing Centre, JLSF 4, A132 (2018)