A performance optimization support framework for GPU-based traffic simulations with negotiating agents

Abstract

To realize a simulation which can handle hundreds of thousands of negotiating agents keeping their detailed behaviors, massive amount of computational power is required. Also having good programmability of agents’ codes to realize complex behaviors is essential to realize it. On deploying such negotiating agents on an agent simulation, it is important to be able to handle detailed behaviors of them, as well as having a large scale simulation to cover important phenomenon that should be observed. There are strong demands to utilize GPU-based computing resources to handle large-scale but very detailed simulations. However, it is not easy task for developers to configure the sufficient parameters to be set on its compilation or execution time, analyzing their performance characteristics on various execution settings. In this paper, we present a framework to assist the coding process of negotiating agents on a traffic simulation, as well as its parameter tuning process on GPU-based programming for simulation developers to utilize GPGPU-based many parallel cores in their simulation programs efficiently. We show how our implemented prototype framework helps simulation developers optimize various parameters and coding-level optimizations to be run on various hardware and software settings.