Model Evaluation of PAT: A Comprehensive Study

Performance of data processing is without doubt the most important criterion in DBMS design. However, the increasing recurring energy cost gradually rivals the benefits of chasing after performance. Therefore, there are strong economic incentives to minimize power consumption within the performance budget, so that the energy cost could be best amortized. Such a goal is difficult to achieve in practice because the power cost consumption tends to vary significantly with actual database workloads. Effective power management solutions are needed to lower the power cost of database servers while maintaining desired performance. Many modern hardware systems provide multiple modes with different power/performance tradeoffs. Among them, changing CPU frequency (and voltage) has been found to be a promising direction towards high energy efficiency in database systems. However, no efforts have been reported on dynamically changing CPU frequency for minimizing energy consumption under a performance bound. In this paper, we present Power-Aware Throughput Control (PAT), a software framework within DBMS that reduces the power consumption of database systems by implementing a feedback PI controller based on workload and system models on power and throughput. The design of PAT based on rigorous control-theoretic analysis brings performance guarantees that are not possible in heuristicsbased database tuning techniques. A workload classification scheme based on fuzzy logic is also developed to capture key information of the workload and further improve the accuracy of system modeling and thus overall PAT performance. Also, comparing with open-loop CPU frequency control, ad hoc control and conventional feedback control, our approach achieves more energy savings regardless of the uncertainties of workload dynamics. We believe this work contributes to a solid foundation of the design and implementation of energy-aware DBMSs.