An Experimental Tool for Elasticity Management through Prediction Mechanisms

摘要

Companies nowadays are widely adopting cloud computing. One important problem encountered to efficiently adopt public IaaS clouds is wasting leased computational resources, which remain active if no policy is established to implement elasticity according to the demand. As a consequence, the cloud use becomes more expensive than expected, leading to higher costs that may turn the company business unattractive for consumers. In this paper, we study this problem in a real company that currently provides resources for its peak demand, which occurs about only 8 hours in weekdays, in order to avoid service disruption. We implemented an experimental resource manager that, through a load prediction mechanism, identifies when the demand will surpass provisioning or when there will be over-provisioning. This enables automatic scaling (up and down) of the infrastructure as needed. We illustrate the application of the system using real data, concluding that through a well trained predicting system and with an intelligent resources manager, it is possible to reduce the waste of computational resources and, consequently, the infrastructure costs.