Power-Aware Throughput Control for Database Management Systems

Performance has been traditionally regarded as the most important design goal for database management systems (DBMSs). However, in recent years, the increasing energy cost gradually rivals the benefit of chasing after performance. Therefore, there are strong financial incentives to minimize power consumption of a database system while maintaining its desired performance, so that the energy cost can be best amortized. Such a goal is challenging in practice because the power consumption of a database system varies significantly with the environment and workloads. Many modern hardware provide multiple modes with different power/performance tradeoffs. However, existing research has not used these power modes sufficiently to achieve the best tradeoff for database services due to the lack of the knowledge on database behavior under different power modes. In this paper, we present Power-Aware Throughput control (PAT), an online feedback control framework for energy conservation at the DBMS level. In contrast to heuristic-based tuning techniques commonly used in database systems, the design of PAT is based on rigorous control-theoretic analysis for guaranteed control accuracy and system stability. We implement PAT as an integrated component of the PostgreSQL system and evaluate it with workloads generated from various database benchmarks. The results show that PAT achieves up to 51.3% additional energy savings despite runtime workload dynamics and model errors, as compared to other competing methods.