The 2nd Joint International Conference on Multibody System Dynamics May 29-June 1, 2012, Stuttgart, Germany Multibody domain decomposition for parallel processing: a wave-based approach to handling interface dynamics C. Smoothey, W. O'Connor UCD School of Mechanical & Materials Engineering University College Dublin, Belfield, Dublin 4, Ireland craig@smoothey.org william.oconnor@ucd.ie Abstract For many good reasons there is growing interest in ways to allow parallel processing of multibody dynamics problems. Some recent approaches include “Domain Decomposition” [1] and “Divide and Conquer” [4]. This paper explores a new approach, reported as work in progress, with initial, promising results. The strategy is an extension of work done on wave analysis of lumped systems in another context [3]. In the approach, a larger system is subdivided into smaller subsystems, which are solved in parallel. Interconnection points are boundaries for each. Dynamic coupling across boundaries is handled in terms of transmitted and reflected motion components (or "waves"), in both directions, across the boundaries. The first test case was to model a 1-D system of an undamped, uniform mass-spring string as if divided into two distinct strings (Fig.1). The uniform dynamics across the imagined boundary imply minimal impedance discontinuity, maximizing the challenge of getting the transfer right. The central mass of the original system was taken as the boundary element, common to the two subsystems. Each subsystem saw this boundary as “active”, like an actuator (see Fig. 1), partly reflecting and partly absorbing incident motion. Wave analyzers [4] were used to determine the right travelling wave in the left system and the left travelling wave in the right system. The sum of these two waves was used to drive the actuators that represent the central mass in the two subsystems. These actuators were then, in effect, partly absorbing the wave in one subsystem and simultaneously launching it into the other subsystem, and partly reflecting. Figure 2: The left travelling wave analyzer. Figure 1: The original and split systems. When implemented, the motion of the split uniform system was found to represent the motion of the original uniform system as accurately as desired. The long-term accuracy was limited only by the order of the wave transfer function, G, utilized in the wave analyzers. The structure of the left travelling wave analyzer is shown in Fig.2. The only data required by the control system is the position of the actuator and of its nearest mass. The wave transfer function, G, estimates the motion transfer between adjacent masses in a system imagined to extend forward to infinity. It can be approximated quite well by second order systems, where the effects of the extension to infinity are modeled simply by a viscous damper. This gives sufficiently good results for many purposes. For perfect results, for example to reproduce very accurately the response of the original system to an impulse continuing for a long time, higher order models of G [3] were used. Usually a system of interest will not be uniform. The division of the larger system into subsystems can then be chosen so that boundaries occur where there is an impedance discontinuity. The uniform system was modified to be “bi-uniform”, with different parameters on each side of the boundary. The left half of the system was configured with k=1, m=1. The right half of the system was configured with k=4, m=1. The wave transfer function GL2R moving across the discontinuity from left hand to right was L X i ( s) GL2R 2 2 X i 1 ( s) s L 2R 2R G R ( s) 2 (1) where GR was the wave transfer function of a hypothetical infinite uniform system configured as in the right-half of the discontinuous system and ωL, ωR are the √(k/m) values in the left, right systems. A symmetrical GR2L wave transfer function was also constructed for the opposite direction. The control system was the same as in Fig.1 and Fig.2 except that two of the four transfer functions changed. A simulation was conducted with a mass-spring string containing 41 masses. A unit impulse was launched into the left side of the system. Figures 3, 4 show the mass displacements of the original and split systems after 180 seconds. The centre black line is the interface point. The maximum absolute error of the bi-uniform, split system was 351 nanometres where the maximum absolute displacement is 195 millimetres. The results were obtained using 11th order G estimates. In practice such high, longterm accuracy would not be required so lower order G estimates would more than suffice for most engineering purposes, especially when damping is present. Figure 3: Original bi-uniform system after 180 seconds simulation time. Figure 4: Split bi-uniform system after 180 seconds simulation time. A similar approach was then tested for a completely non-uniform cascaded mass-spring system, again split in the middle. The results were remarkably similar to the bi-uniform case, with arbitrarily high accuracy achievable. Work is now in progress on a flexing system of n-links with torsional springs at the joints. Detailed results will be available in time for the full paper. The background ideas for waveresolving in such flexural systems have already been successfully implemented in other applications. The proposed method promises various benefits for parallel processing. Obviously each subsystem would be solved simultaneously by different processor cores, utilizing wave-based coupling between the subsystems. These subsystems would clearly be significantly smaller and more manageable than the whole. The computational overhead for wave resolution is small. Furthermore, because the interface “wave” variables are components of the motion of the local systems, it is not necessary to solve explicitly for the action-reaction forces and torques at the subsystem interfaces. Thus, as in analytical mechanics in general (vs. vector mechanics), the internal, constraint forces are handled automatically and need not be determined explicitly in the solution scheme. This is a significant extra benefit of the method. References [1] Imanishi, E.; Nanjo, T.: “Fast Simulation Of Flexible Multibody Dynamics Using Improved Domain Decomposition Technique”, Proceedings of the ECCOMAS Thematic Conference on Multibody Dynamics 2011, July 4-7, 2011, Brussels. [2] McKeown, D.J.: “Wave Based Control of Flexible Mechanical Systems” (Doctoral Dissertation, University College Dublin). Dublin, 2009. [3] O'Connor, W.J.: “Wave-Echo Control of Lumped Flexible Systems”, Journal of Sound and Vibration, Vol. 298, pp. 1001-1018, 2006. [4] Poursina, M.; Anderson K.S.: “Constant Temperature Simulation Of Articulated Polymers Using Divide-And-Conquer Algorithm”, Proceedings of the ECCOMAS Thematic Conference on Multibody Dynamics, 2011, July 4-7, 2011, Brussels.