Dynamic domain expansion in smoke spread simulations with ARTSS: Speedup and overhead

Abstract This paper describes the impact and consequences of a dynamic domain expansion in a smoke simulation performed in the software ARTSS. This software is developed with the aim to conduct real-time or even prognosis computations by using GPUs as the main computational architecture. Further runtime acceleration is proposed by means of a dynamic extension of the computational domain. This approach is based on the reduction of the computational domain, which is dynamically adopted to calculate only the domain of interest, e.g. regions containing smoke. Here, the domain starts as a localised region and is extended based on prescribed criteria. This contribution outlines the initial implementation. However, to understand the impact of an expansion, the overhead caused by the expansion process, the influence on the numerical result and on the runtime, as well as the used expansion parameters, are investigated. In general, an increased acceleration can be eventually observed at the costs of accuracy due to the reduced domain. The overhead and accuracy can be controlled by the method's parameters. The loss of accuracy depends strongly on which expansion methods and setting are used. With more complex expansion methods, the loss of accuracy can be reduced.