Gridlock resolution in a GPU-accelerated traffic queue model

While agent-based mobility simulations are becoming more widely used within the research community, the complexity of applied agent-based models and the scale of scenarios are growing. To tackle the increased computational demand, hardware accelerators such as GPUs are used for faster execution of large-scale scenarios. However, quite often, specialized hardware accelerators have limited capabilities for memory management and code execution. Such limitations pose an additional challenge to the implementation of traffic simulation models. This paper focuses on a gridlock situation that is typical for spatial traffic queue models and which has to be resolved in order to run simulations adequately. We propose and evaluate three different strategies for gridlock resolution in a mesoscopic traffic queue model implemented on GPU. The results show that, while the final simulation outcomes are similar for each strategy, the least realistic strategy provides faster convergence at a lower memory footprint, whereas the most realistic strategy is more suitable for visualization at the expense of slower convergence and up to 25% increase of the run time.