Database-Managed CPU Performance Scaling for Improved Energy Efficiency

Dynamic voltage and frequency scaling (DVFS) is a technique for adjusting the speed and power consumption of processors, allowing performance to be traded for reduced power consumption. Since CPUs are typically the largest consumers of power in modern servers, DVFS can have a significant impact on overall server power consumption. Modern operating systems include DVFS governors, which interact with the processor to manage performance and power consumption according to some system-level policy. In this paper, we argue that for database servers, DVFS can be managed more effectively by the database management system. We present a power-aware database request scheduling algorithm called POLARIS. Unlike operating system governors, POLARIS is aware of database units of work and database performance targets, and can achieve a better power/performance tradeoff by exploiting this knowledge. We implemented POLARIS in SHORE-MT, and we show that it can improve both power consumption and performance relative to operating system baselines.