Towards billion-scale social simulations

Abstract

Many social simulations can be represented using mobile-agent-based model in which agents moving around on a given space such as evacuations, traffic flow and epidemics. Whole planet simulation with billions of agents at microscopic level helps mitigate the global crisis. It introduces new technical challenges such as processing and migrating many agents and load balancing among hundreds of machines. To overcome these challenges, well-designed software architecture of a simulator is essential. In this research, we proposed agent-based complex cellular automata architecture (ABCCA) and studied the performance and scalability of two cell-based processing models, through simple traffic flow simulation on multi-core distributed system. The experiments show that the computation speedup can be achieved by reducing granularity of tasks and processing only active spaces. We achieved running the traffic flow simulation with one billion of agents in almost real time on 1,536 CPU cores of total 128 machines of TSUBAME supercomputer.