Parallel implementation of domain decomposition algorithm for molecular dynamics

The objective of this study is to apply a domain decomposition algorithm to the La- grangian based Discrete Particle Method (DPM). The latter deals with the thermal decomposition of solid e.g. biomass fuel particles. It considers each particle as an individual entity that is represented by an instantiation of a class. Modelizing thermal conversion of biomass in real systems involves a large number of particles. This can be achieved in a reasonable computing time only through a parallel implementation able to distribute the particles e.g. objects onto the participating processors. This pa- per present such an implementation based on Orthogonal Recursive Bisection (ORB) method. Due to the fact that the particles may take arbitrary positions within the do- main, a particular issue addressed by the domain decomposition technic used in this work is to generate a load balance for each processors as uniform as possible. Fur- thermore, the particles are coupled via heat transfer. One challenge for the designed algorithm is then to identify the nearest neighbours of each particle so that the nec- essary information can be communicated between them. Since the positions of all particles are subject to change and may migrate from one processor to another, the communication links together with the number of neighbors are highly dynamic. The implementation is carried out using the KAAPI API, a C++ library for parallel pro- gramming that allows to execute ?ne/medium grain multithreaded computation with dynamic data ?ow synchronizations. First results are very promising since they indi- cate that our algorithm creates sub domains with an average imbalance ranging from 2.5% to 6.3% for uniformly distributed particles.