Predictive approaches to cloud computing costs reduction

Cloud Computing has become more and more affordable, it has seen an important growth in terms of adoption among all the IT companies in the last few years, overtaking the previously used on-premise solutions. Indeed, Cloud Computing is generally cheaper and easier to set-up with respect to on-premise solutions, requiring no maintenance and coming with more service stability, generally higher up-time and dedicated support from experienced companies like Google or Amazon in case of problems. The easiness of Cloud Computing, however, may lead developers to misuse computational resources mainly because it is very easy to add more memory or more processing units to any application. Moreover, developers never know the precise amount of resources their applications will need, thus exceeding with resources allocation just to be ready to handle all possible worst cases. In this work we will address the misuse of cluster resources from two perspectives, focusing our attention on batch jobs. Firstly, we will address the problem of scheduling several jobs into the same clusters, thus implementing resources sharing among jobs. Then, we will focus on automatically resizing cluster's resources (expressed in terms of computational machines) in order to always provide the best fit between cluster's resources and user's jobs. Moreover, we will prove that better scheduling decisions can be taken by exploiting the prior knowledge we have on our jobs in order to build Machine Learning models. In conclusion, we will show how we managed to obtain a 65% saving in costs of cloud computing in our team by adopting our system as an intermediate layer between the developers and the Cloud Computing platform.