Dynamic management of resources and workloads for RDBMS in cloud: a control-theoretic approach

As cloud computing environments become explosively popular, dealing with unpredictable changes, uncertainties, and disturbances in both systems and environments turns out to be one of the major challenges facing the concurrent computing industry. My research goal is to dynamically manage resources and workloads for RDBMS in cloud computing environments in order to achieve ``better performance but lower cost", i.e., better service level compliance but lower consumption of virtualized computing resource(s). Nowadays, although control theory offers a principled way to deal with the challenge based on feedback mechanisms, a controller is typically designed based on the system designer's domain knowledge and intuition instead of the behavior of the system being controlled. My research approach is based on the essence of control theory but transcends state-of-the-art control-theoretic approaches by leveraging interdisciplinary areas, especially from machine learning. While machine learning is often viewed merely as a toolbox that can be deployed for many data-centric problems, my research makes efforts to incorporate machine learning as a full-fledged engineering discipline into control-theoretic approaches for realizing my research goal. My PhD thesis work implements two solid systems by leveraging machine learning techniques, namely, ActiveSLA and SmartSLA. ActiveSLA is an automatic controller featuring risk assessment admission control to obtain the most profitable service-level compliance. SmartSLA is an automatic controller featuring cost-sensitive adaptation to achieve the lowest total cost. The experimental results show that both of the two systems outperform the state-of-the-art methods.