FLAME GPU 2: A framework for flexible and performant agent based simulation on GPUs

Agent based modelling (ABM) offers a powerful abstraction for scientific study in broad range of domains. The use agent simulators encourages good software engineering design such as separation concerns, that is, the uncoupling model description from its implementation detail. A major limitation current approaches to ABM simulation is trade off between simulator flexibility and performance. It common highly optimised simulations, those which target graphics processing units (GPU) hardware, are implemented standalone software. This work presents framework (FLAME GPU 2) balances with performance general purpose ABM. Methods ensuring high computational efficacy demonstrated by, minimising data movement, device utilisation by exploiting opportunities concurrent code execution within through ensembles simulations. novel hierarchical sub-modelling approach also presented can be used certain types recursive behaviours. feature shown essential providing mechanism resolve competition resources agents parallel environment would otherwise introduce race conditions. To understand characteristics software, benchmark millions explore parametrically investigate model. Performance speedups 3.5 × $$ \times 10 respectively over baseline implementation. Our demonstrate algorithm movement occurs result desire simultaneously occupy discrete areas ‘resource’. implement classical socio-economics model, Sugarscape, populations up 16M agents.