Efficient Parallel Simulation over Social Contact Network with Skewed Degree Distribution

Social contact network (SCN) models the contacts between people by their daily activities. It can be formalized by an agent-to-location bipartite graph. The simulations over SCN are employed to study the complex social dynamics such as information propagation and disease spread among large-scale population. A challenge to the simulation is the skewed degree distribution of SCN, which contains a few hub locations with large numbers of visitors. The skewed degree distribution can cause load imbalance for parallel simulation and greatly degrade the execution performance. This paper proposes an approach which decomposes hub locations into small splits. Thus, SCN can be partitioned with better balanced workloads and multiple splits are able to run in parallel. Based on the pattern of information transmission between agents, we duplicate necessary data among splits to ensure the correctness of simulation. Furthermore, we enhance the parallel algorithm of SCN simulation to reduce the additional overhead from communication between splits. Finally, we build an experiment with epidemic simulation on an open dataset. The experimental results demonstrate that our approach achieves 14~35% performance improvement compared with the partitioning method without decomposition of hub locations.