A computing resources prediction approach based on ensemble learning for complex system simulation in cloud environment

Abstract

Cloud computing provides a new infrastructure for the research of complex system simulation (CSS). However, insufficient computing resource allocation results in lower performance of a CSS application. On the other hand, excessive computing resource allocation will lead to the increase of simulation communication overhead and simulation synchronous computing. Therefore, accurate computing resource prediction is important to achieve optimal scheduling for CSS applications in the cloud environment. In this paper, a computing resource prediction approach based on ensemble learning has been proposed, which includes a cloud computing resource prediction framework and an intelligent ensemble algorithm. The framework with three–level architecture (simulation as a service, cloud computing resource predictor, and cloud computing resource pool) can provide computing resources to deal with the demands of the simulation applications. The intelligent ensemble algorithm uses an Accuracy and Relative Error-based Pruning algorithm to ensure the effective ensemble of base models (support vector machine, decision tree, and k-nearest neighbor). To improve the performance of the intelligent ensemble algorithm, a Feature Capability-based forward search Feature Selection algorithm is introduced to reduce redundancy between features. The experiments are presented to demonstrate that the intelligent ensemble algorithm can achieve higher accuracy by 4%-20% when compared with existing resource prediction models such as Regressive Ensemble Approach for Prediction, Bayesian, Linear Regression, Random Forest, and Fuzzy Neural Network.