Diffusion based adaptive load-balancing for domain decomposition in particle simulations

Particle simulations, such as molecular dynamics, dissipative particle dynamics or Brownian dynamics are nowadays standard methods for studying the structure and dynamics of complex systems in various geometries. Knowing the interaction potentials between particles it is possible to integrate the governing equations of motion and to study the phaseor configurationspace of the given system. Studying larger systems on longer time scales induces a strong demand for the parallelization of the underlying algorithms and various simulation packages exist, which show good parallel scaling for homogenously distributed systems. For MD simulations, applying domain decomposition schemes, it was shown that a scaling up to > 100 k compute cores is possible. The situation gets problematic, when the particle systems exhibit inhomogenous particle distributions, e.g. when simulating clusters, self aggregation, phase seperation or supercritical systems, to name a few. If a uniform partitioning of geometrical domains is assumed for the parallel decomposition in these systems, the work load W (p) on different processors will be non-homogenous, since W (p) = ∫ Ωp dr ρW (r) strongly depends on the work density ρW (r). This density might be differently defined, but a common measure would be the number of particles on each processor or the time, which is spent to calculate mutual interactions between particles. In the case where the work is inhomogenously distributed, the parallel scaling of the underlying problem is determined by the slowest process. A