The impact of minimising data movement on the overall performance of the simulation of complex systems applied to FLAME GPU

GPUs have been demonstrated to be highly effective at improving the performance of Multi-Agent Systems (MAS). One of the significant limitations of further performance improvements is in the memory bandwidth required to move agent data through the GPU’s memory hierarchy. This thesis investigates the impact of data dependency on the FLAME GPU framework’s overall performance as an example of Agent Based Modelling (ABM) platforms. This investigation includes discovering data dependency within FLAME GPU models. Two methods are proposed in order to minimise data movement during simulation using dependency information: (i) a functional method which is based on the concept of merging and splitting agent function; and (ii) data-aware method which uses of data dependency information to access a subset of agent and message memory at the variable level. This thesis also develops a method that allows automatic discovery of data dependencies from existing FLAME GPU models. This method is based on parsing an agent function file of a FLAME GPU model to extract all agent functions’ data dependencies. The scalability, computational complexity, internal memory requirements, and homogeneity of the agent and population of the model are examples of factors that may affect ABM applications’ overall performance. This thesis presents a standard benchmark model designed to observe the system behaviour while testing these factors. An evaluation of the performance impact of minimising data movement has been carried out by implementing the proposed FLAME GPU methods using the benchmark model and the number of existing FLAME GPU models. The comparison between the current and new system shows that reducing data movement within a simulation improves overall performance.