A simulation task partition method based on cloud computing resource prediction using ensemble learning

Abstract

The calculation processing and interactions between complex system simulation entities in a cloud-computing environment exhibit dynamic change. However, parallelized task partitioning methods often ignore these dynamic characteristics, which makes it difficult to improve execution efficiency for complex system simulations. Therefore, we propose a two-stage simulation task partition method that computes the resource prediction information. In the first stage, the features of collected data are analyzed to select the optimal feature dimensions. A stacking ensemble learning method is used to predict the simulation runtime under the condition of allocating various computing resources, and the number of resources required for the shortest runtime is obtained by sorting the predicted results. In the second stage, a multi-weight graph structure is designed to represent the dynamic interaction of simulation entities, with the goal of minimizing the load imbalance on each computing node, as well as the maximum number of communications between nodes. Subsequently, the simulated annealing optimization algorithm is used to partition multiple weight graphs and map simulation tasks to resources. We compared our multi-weight graph partition method with the Metis and Chacos models. The experimental results demonstrate that, while all three methods can achieve the shortest runtime with the predicted number of computing resources, our algorithm can further reduce the runtime by up to 26%.