Assessment and optimization of MCNP memory management for detailed geometry of nuclear fusion facilities

Abstract The main nuclear fusion facilities world-wide rely on MCNP to conduct the nuclear analysis activities. It is a computational tool based on Monte Carlo method and probably the most reliable one to solve neutral particles radiation transport problems. The increasing availability of High Performance Computing infrastructure has stimulated the development of tools to translate complex CAD model to MCNP, helping to reduce uncertainties associated with geometry modelling. However, an associated need of RAM memory has emerged, reaching the limits of memory available per CPU in current parallel super computers. Currently, some ITER analyses use models requiring more memory per CPU than available in super computers like Marconi. In these cases the simulations need to reduce the number of CPU used per node, leading to a waste of computational resources. This paper assesses the MCNP memory management related to the geometry and proposes some hints to reduce it acting on the MCNP models. This same study identified the origin of the problem being a coding in MCNP leading to an overestimation of the memory allocation. Modifications in MCNP have been implemented to reduce the memory consumption; these changes have improved noticeably the code performances.