Simulation runtime prediction approach based on stacking ensemble learning

Abstract

Many technologies in complex system simulation (CSM), such as resource scheduling and load balancing, largely rely on historical data with different characteristics to predict the future. The accuracy of runtime prediction has a significant impact on scheduling performance. However, with cloud computing becoming the main infrastructure for deploying CSM applications, the current prediction methods are difficult to adapt to the dynamic changes of cloud computing resources. Insufficient computing resource allocation will be difficult to support the efficient operation of simulation. In addition, excessive computing resource allocation will lead to higher computing and data communication costs. Therefore, a simulation runtime prediction approach based on stacking ensemble learning has been proposed, which uses the characteristic variables of simulation applications (such as the number of simulation entities, the number of simulation events, the simulation time, etc.) and the performance monitoring data of computing resources (such as CPU utilization, memory utilization, etc.) as the characteristic inputs. The machine learning algorithms such as XBG, SVG, MLP are integrated by stacking model, and the performance of the integrated learning algorithm is comprehensively evaluated by mean absolute error (MAE), accuracy (ACC), root mean square error (RMSE) and coefficient of determination (R2). Experimental results show that the proposed algorithm improve the prediction accuracy by 3% - 24% when compared with existing machine learning prediction methods.