Evaluating Sensitivity of Auto-Scaling Decisions in an Environment with Different Workload Patterns

Cost-performance trade off is one of the critical challenges in cloud computing environments. Predictive auto-scaling systems mitigate this issue by scaling in/out system automatically based on performance prediction results. The goal of this research is to investigate the impact of different prediction results on the scaling actions generated by predictive auto-scaling systems. In this study, predictive auto-scaling systems are divided into Predictor and Decision-maker components, and experiments have been conducted to measure the influence of the Predictor results on the Decision-maker output. We have used Support Vector Machine (SVM) and Neural Networks (NN) as the Predictor component and the threshold-based technique as the Decision-maker component. In addition, the influence of different workload patterns on the prediction accuracy of SVM and NN has been investigated in this paper. The experiment results show that although either of the prediction algorithms (i.e., SVM and NN) is suitable for a specific workload pattern, SVM and NN predictions lead to similar scaling actions in 97.9% of time, which suggests that focus should be on the decision-maker techniques