Measuring prediction sensitivity of a cloud auto-scaling system
Elasticity is one of the key benefits of cloud computing which helps customers reduce the cost. Although elasticity is beneficiary in terms of cost, obligation of maintaining Service Level Agreements leads to necessity in dealing with the cost-performance trade-off. Proactive auto-scaling is an efficient approach to overcome this problem. In this approach scaling actions are generated based on prediction results. Recently, several research studies have been focusing on improving prediction accuracy in order to improve the efficiency of auto-scaling mechanisms. However, the sensitivity of auto-scaling mechanisms to the prediction results is neglected in the domain. In this work we have investigated the sensitivity of auto-scaling mechanisms to the prediction results by evaluating the influence of performance predictions accuracy on the auto-scaling actions. Specifically, we have compared actions of threshold based scaling techniques which are generated based on Support Vector Machine (SVM) and Neural Networks (NN) predictions. Our experimental results show that although SVM is more accurate than NN, scaling decisions made by the two algorithms are identical in 91.5% of the time. Furthermore, we have shown that the optimal training duration for SVM and NN is about 60% of experiment duration.
|Keywords||Cloud computing, Machine learning, Performance prediction, Resource provisioning|
|Conference||38th Annual IEEE Computer Software and Applications Conference Workshops, COMPSACW 2014|
Nikravesh, A.Y. (Ali Yadavar), Ajila, S, & Lung, C.H. (2014). Measuring prediction sensitivity of a cloud auto-scaling system. Presented at the 38th Annual IEEE Computer Software and Applications Conference Workshops, COMPSACW 2014. doi:10.1109/COMPSACW.2014.116