Dynamic Provisioning and Run-time Optimization of Cloud Workloads.

Cloud computing has become very important nowadays for companies, and many of them started offloading job computations instead of increasing the on-premise capacity: the dispatch of those jobs is usually done manually by users, leaving them the choice of the instance and provider to be used. The scope of this thesis is to analyze possible improvements given by the introduction of machine learning in the decision process: the idea was to create a new unit, independent from the scheduler already implemented in the company, having the possibility to extend it to every system deployed. The user is still the author of the choice because he/she is given more information to improve the decision and is not bypassed. The designed system consists of 2 internal predictors, so it assumes the name of two-stage predictor. Each stage exploits machine learning, but instead of using just one single algorithm, it uses several algorithms to achieve better performance: each one is independently tuned with a set of hyperparameters, with the training on available data as the following step. After the training phase, the best performing algorithm is selected and used for that predictor, in a way that each predictor is independent of the algorithm that is used. The system was tested on 2 projects, considering the combination of two features: data augmentation and Continuous Machine Learning (CML). The analysis conducted shows two important and different outcomes, that are different but bring to the same final considerations. The first one shows how the system can be wrong if not constantly trained when new data is available, leading to a higher cost with respect to the optimal solution. The second one shows that the system, if well trained and updated, can follow the evolution over the time and learn from data, leading to a potential saving of about 10%. Different results could be obtained under different initial conditions. As concluding remark, the introduction of machine learning to the job dispatch problem can be effective and lead to optimal solutions only under some conditions. The advantage is the total cost reduction if the current dispatching is not optimal or validate the latter if there is no reduction. The drawback is that the system must be kept updated to follow the evolution of data through time, otherwise, prediction can become inaccurate and lead to much higher costs than expected: the conditions are the number of data available and the application of CML.