Optimizing the coupling in parallel air quality model systems

Today, parallel computers facilitate complex simulations of physical and chemical processes. To obtain more accurate results and to include multiple aspects of environmental processes, model codes of different scientific areas are coupled. An often used coupling strategy is to run the individual codes concurrently on disjoint sets of processors, as this keeps the codes mostly independent. However, it is important to improve the workload balance between the codes to achieve a high efficiency on parallel computers. In this paper, the parallel air quality model system LM-MUSCAT is presented. It consists of the chemistry-transport model MUSCAT and the meteorological model LM. Since an adaptive time step control is applied in MUSCAT the overall load fluctuates during runtime, especially at applications with highly dynamical behavior of the simulated processes. This causes load imbalances between both models and, consequently, an inefficient usage of the parallel computer. Therefore, an alternative coupling method is investigated. In this approach, all processors calculate alternately both models, whereby the load is distributed equally. Performance tests show that this ''sequential'' approach is well suited to increase the efficiency of coupled systems that have workload fluctuations in one or more models. In general, load variations can occur in models which use adaptive grid techniques or an adaptive step size control. Systems using such techniques can take benefit from the described coupling approach.