Previsão de acesso de objetos em serviços de armazenamento em nuvem em camadas

Abstract

The exponential growth of global data poses significant challenges to cost management in cloud storage infrastructures. Cloud providers offer tiered storage systems with different cost and performance classes, creating a fundamental trade- off between financial savings and access latency. Manual data allocation to these tiers is impractical at scale, motivating the development of intelligent automatic optimization systems. This thesis investigates the application of machine learning techniques for tiered storage optimization through data access pattern prediction. The central hypothesis proposes that predictive approaches outperform traditional strategies (static and reactive) in cost efficiency and quality of service. This hypothesis is operationalized through three sub-hypotheses: (1) the feasibility of predicting future access patterns with sufficient accuracy for proactive decisions; (2) the effectiveness of a control mechanism allowing administrators to express cost-latency preferences; and (3) the gains provided by extending from binary to multiclass classification. The research develops in three evolutionary stages. Initially, a binary classification system (Hot/Cold) demonstrated savings of up to 37% compared to baseline strategies. Subsequently, the integration of user preferences through the cost-latency weight parameter (wc) enabled explicit control over the trade-off. Finally, the extension to multiclass classification (Cold/Warm/Hot) achieved savings of up to 40%, revealing that machine learning models adapt significantly better than reactive strategies to the three-tier scenario. The methodology is based on a comparative analysis of 13 algorithms, including Random Forest, Support Vector Machines, k-Nearest Neighbors, and ensemble methods. The evaluation uses real production traces from Dropbox, covering two points of presence with distinct characteristics. Beyond traditional classification metrics, the research employs the Relative Cost Savings (RCS) metric, which incorporates the differentiated costs of storage classes. Results demonstrate that the proposed multiclass framework consistently outperforms baseline strategies, achieving accuracy above 80% and cost savings of up to 40%. Pareto frontier analysis reveals that different algorithms are optimal in different regions of the cost-latency spectrum, transforming the framework into a decision support tool. Validation with real data confirms the practical applicability of the solution. The main contributions include: (1) an evolutionary framework from binary to multiclass classification with production validation; (2) the wc parameter as an original mechanism for integrating user preferences; (3) comprehensive comparative analysis of algorithms for the storage optimization domain; and (4) demonstration that the empirical Pareto frontier is composed of multiple algorithms. The work establishes a solid foundation for future research in intelligent storage resource optimization. Keywords: Machine Learning, Cloud Storage, Cost Optimization, Tiered Storage, Access Pattern Prediction, Multiclass Classification, Multi-objective Optimization