Dynamically Adaptive Partition-Based Data Distribution Management

Abstract

Performance and scalability of distributed simulations depends primarily on the effectiveness of the employed data distribution management (DDM) algorithm, which aims at reducing the overall computational and messaging effort on the shared data to a necessary minimum. Existing DDM approaches, which are variations and combinations of two basic techniques, namely region-based and grid-based techniques, perform purely in the presence of load differences. We introduce the partition-based technique that allows for variable-size partitioning shared data. Based on this technique, a novel DDM algorithm is introduced that is dynamically adaptive to cluster formations in the shared data as well as in the physical location of the simulation objects. Since the re-distribution is sensitive to inter-relationships between shared data and simulation objects, a balanced constellation has the additional advantage to be of minimal messaging effort. Furthermore, dynamic system scalability is facilitated, as bottlenecks are avoided.