An efficient dynamic load balancing method for simulation of variable structure systems

Multi-core or many-core machines become increasingly pervasive. Parallel simulation can achieve high efficiency by employing multi-threaded parallel execution of models and fast inter-thread communication between models. However, load imbalance among cores degrades performance of the simulation system greatly. This is especially true for simulation of variable structure systems, due to dynamic addition or removal of models. An efficient load balancing method under multi-cores or many-core environments has been proposed, based on our threaded parallel simulator-Ivy [1]. The method employs shared models and shared interaction structure in one OS process to achieve real dynamic migration of loads among cores, with little intervention to the simulation execution, while the copying and transferring of model state are avoid. The proposed metrics for load migration take variation of the system composition into consideration and can be applicable to both dedicated and non-dedicated multi-core machines. The metrics can reflect the computation load of a fine-grained model by ratio of the overall CPU time (consumed by events processing) to virtual advance time, while communication cost is properly considered. The experiment results show that our method can improve efficiency of both common simulation and variable structure simulation evidently.