Section: New Results

Dynamic Load Balancing of Monte Carlo Particle Transport Applications

Participants : Thomas Gonçalves, Marc Pérache [CEA DAM, Bruyères le Châtel] , Frederic Desprez, Jean-Francois Mehaut.

Monte Carlo particle transport applications consist in studying the behavior of particles moving about a simulation domain. Particle distribution among simulation domains is not uniform and changes dynamically during simulation. The parallelization of this kind of applications on massively parallel architectures leads to solve complex issues of workloads and data balancing among numerous compute cores.

This research work started by identifying parallelization pitfalls of Monte Carlo particle transport applications using theoretical and experimental analysis of reference parallelization methods. A semi-dynamic load-balancing based on partitioning techniques has then been proposed. Finally, we designed a dynamic approach which is able to redistribute workloads and data while keeping a low communication volume. Compared to the perfectly balanced domain replication method using strong scaling measurement, the dynamic approach leads both to speedups and reduction of memory footprint.

This work is part of the Thomas Gonçalves's PhD thesis defended in September 2017 at TERATEC (Bruyères le Châtel). The main contributions of this work were also presented in the ParCo conference [14] .