In the context of a Monte Carlo calculation for the evolution of a PWR (pressurized water reactor) core, it is necessary to compute a very large number of neutron-nucleus reaction rates, involving a data volume that can exceed the memory capacity of a compute node on current supercomputers. Within the Tripoli-5 framework, distributed memory architectures have been identified as targets for high-performance computing deployment. To leverage such architectures, data decomposition approaches must be used, particularly for reaction rates. However, with a classical parallelization method, processes have no particular affinity for the rates they host locally; on the contrary, each rate receives contributions uniformly from all processes. Access to decomposed data can be costly when it requires intensive use of communications. Nevertheless, one-sided communication mechanisms, such as MPI RMA (Message Passing Interface, Remote Memory Access), make these accesses easier both in terms of expression and performance.
The objective of this thesis is to propose a method for partial data decomposition relying on one-sided communication mechanisms to access remotely stored data, such as reaction rates. Such an approach will significantly reduce the volume of data stored in memory on each compute node without causing a significant degradation in performance.