Efficient Parallelization of Multibody Systems Incorporating Co-Simulation Techniques
Co-simulation methods can be used advantageously in the field of multi-disciplinary simulations. Another applicability of co-simulation methods is the parallelization of large monodisciplinary dynamical models. This work focuses on the reduction of computation time that can be achieved in the simula...
| Summary: | Co-simulation methods can be used advantageously in the field of multi-disciplinary simulations. Another applicability of co-simulation methods is the parallelization of large monodisciplinary dynamical models. This work focuses on the reduction of computation time that can be achieved in the simulation of multibody systems by partitioning a monolithic model into a set of coupled subsystems. The connection between the subsystems can be described in various ways. In this work, different subsystems are coupled by nonlinear constitutive equations (applied force coupling approach). Information (i.e. coupling variables) is only exchanged between the subsystems at distinct communication-time points (macro-time points). Within each macro-time step, the unknown coupling variables are approximated by extrapolation polynomials. The essential point is that the subsystems are integrated independently of each other between the macro-time points. If a Jacobi type co-simulation scheme is used, all subsystems can be solved in parallel.
A main drawback of many co-simulation implementations is that they are based on an equidistant communication-time grid.
Using a constant macro-step size may in many practical applications be not very efficient with respect to computation time, especially in connection with highly nonlinear models or in context with models with strongly varying physical parameters. In this work, explicit and implicit co-simulation approaches which incorporate a macro-step size and order control algorithm, are presented. Numerical examples show the benefit of this implementation and the significant reduction in computation time compared to an implementation with an equidistant communication-time grid. In addition, a comparison between a co-simulation model and a monolithic model demonstrates the great computation time reduction which can be achieved due to the parallelization and the multirate character of the proposed co-simulation methods.
The co-simulation approaches are fully parallelized by a hybrid MPI and OpenMP implementation. The resulting computation time of the implemented co-simulation approaches is analyzed in detail. The influence of various simulation parameters on the computation time is studied and sources of computational overhead are identified. |
|---|