Scalability and composability of flow accumulation algorithms based on asynchronous many-tasks

Models simulating the state of the biological and physical environment can be built using frameworks that contain pre-developed data structures and operations. To achieve good model performance it is important that individual modelling operations perform and scale well. Flow accumulation operations that support the use of criteria for selecting how much material flows downstream are an important part in several Earth surface simulation models. For these operations, no algorithms exist that perform, scale, and compose well. The objective of this study is to develop these algorithms, and evaluate their performance, scalability, and composability. We base our algorithms on the asynchronous many-task approach for parallel and concurrent computations, which avoids the use of synchronization points and supports composability of modelling operations. The relative strong and weak scaling efficiencies when scaling a flow accumulation operation over six CPU cores in a NUMA node are 83% and 84% respectively. The relative strong and weak scaling efficiencies when scaling a case-study model over four cluster nodes are 73% and 84%. Our algorithms are composable: the latency of executing two flow accumulation operations combined is lower than the sum of their individual latencies.