Uma arquitetura autonômica para alocação preditiva de recursos no gerenciamento de capacidade em nuvem utilizando redes neurais

Abstract

Cloud computing has been identified as a solution for the rational use of Information Technology resources. Cloud service providers offer shared environments that can scale to meet their customers' fluctuating requirements. The challenge imposed is the use of mechanisms capable of optimizing the use of resources and, simultaneously, ensuring that the performance of these services continues to meet the Quality of Experience (QoE), Quality of Service (QoS) metrics, as well as the Service Level Indicators (SLI) and the respective Service Level Agreements (SLA) established. Providers need to offer autonomic mechanisms to promote resource scalability promptly, while customers need to trust the performance and costs involved in negotiations. In the Cloud Resource Management process, several predictive resource scheduling approaches have already been proposed to overcome the limitations of conventional reactive techniques. However, such methods have yet to demonstrate satisfactory cost, performance, and autonomy results. This research proposes an Autonomic Cloud Resource Management Model that combines reactive and predictive features for resource scheduling. To support the predictive provisioning of resources, Recurrent Neural Networks (RNNs) were used in the Stacked Long Short-Term Memory architecture, seeking to overcome the results already achieved. The MAPE-K model was adopted in the autonomic approach, using the principles of Autonomic Cloud Computing (ACC). A case study was elaborated using experimental traces to demonstrate the proposal's viability. In evaluating the model's accuracy, a comparison between the classic LSTM network and different configurations of the Stacked LSTM network was used. A prototype was implemented using components of cloud infrastructure simulators to analyze the operational viability. The obtained results demonstrated the feasibility of the proposal, bringing as a benefit the use of Stacked LSTMs in predicting the provisioning of cloud resources. As future work, it is intended to evolve the prototype into an operational tool, in open source, to support small and medium-sized service providers and allow capacity planning in the cloud migration process.