The efficiency of parallel solvers for large multibody systems is affected by the topology of the network of constraints. In the most general setting, that is the case of problems involving contacts between large numbers of parts, the mechanical topology cannot be predicted a priori and also changes during the simulation. Depending on the strategy for splitting the computational workload on the processing units, different types of worst case scenarios can happen. In this paper we discuss a few approaches to the parallelization of multibody solvers, ranging from the fine-grained parallism on GPU to coarse-grained parallelism in clusters, and we show how their bottlenecks are directly related to some graph properties of the mechanical topology. Drawing on the topological analysis of the constraint network and its splitting, lower bounds on the computational complexity of the solver methods are presented, and some guidelines for limiting the worst-case scenarios in parallel algorithms are put forward.

This content is only available via PDF.
You do not currently have access to this content.