Solving Large Multibody Dynamics Problems on the GPU

Abstract

This paper describes an approach for the dynamic simulation of complex computer-aided engineering models where large collections of rigid bodies interact mutually through millions of frictional contacts and bilateral mechanical constraints. Thanks to the massive parallelism available on today's GPU boards, we are able to simulate sand, granular materials, and other complex physical scenarios with one order of magnitude speedup when compared to a sequential CPU-based implementation of the discussed algorithms.

Enabling Computational Dynamics in Distributed Computing Environments Using a Heterogeneous Computing Template

“…Recent granular dynamics results [11] indicate reduction in simulation times on the order of 50-60, and a factor of 60-70 for collision detection of ellipsoids [21]. Discrete Element Method simulation results reported by [10] suggest more than two orders of magnitude speedup when going from sequential to GPU-enabled parallel computing.…”

“…Two order of magnitude reductions in simulation times and increases in problem size are demonstrated when using heterogeneous CPU/GPU computing for collision detection, where problems with up to six billion collision events were solved in less than three minutes. Although not discussed here, heterogeneous computing has motivated research into numerical methods for the parallel solution of large differential variational inequality problems [11], and has also been used very effectively in distributed visualization tasks where postprocessing times for simulation visualization were reduced by more than one order of magnitude [4]. Beyond its immediate relevance in solving many-body dynamics problems, heterogeneous CPU-GPU computing promises to become a computational paradigm that can address stringent efficiency needs in Scientific Computing applications in diverse fields such as climate modeling, quantum chemistry, fluid dynamics, and biochemistry.…”

“…For Discrete Element Method approaches that use explicit numerical time integration [10], DDEP-enabled communication between X and Y occurs once per time step. For Differential Variational Inequality approaches state information should be exchanged using DDEP multiple times during the same time step due to a Cone Complementarity Problem that needs to be iteratively solved to recover the frictional contact force acting between bodies 4 and 5 [11]. Similarly, for approaches using implicit integration, communication must occur multiple times per time step during the iterative solution process.…”

“…Chrono::Engine uses a formulation based on a differential variational inequality (DVI). Compared to a penalty approach (as in DEM), the DVI approach enforces contacts as non-penetration constraints and allows larger integration time steps at the cost of higher computational effort at each step [11].…”

Parallel Computing in Multibody System Dynamics: Why, When, and How

“…Mobility Simulation A detailed account of how GPU computing is used to solve multibody dynamics problems can be found in [54]. The key kernels of the GPU implementation are listed in the following pseudocode:…”