Towards Dynamic Green-Sizing for Database Servers

This paper presents two techniques for reducing the power consumed by database servers. Both techniques are intended primarily for transactional workloads on servers with memory-resident databases. The first technique is databasemanaged dynamic processor voltage and frequency scaling (DVFS). We show that a DBMS can exploit its knowledge of the workload and performance constraints to obtain power savings that are more than twice as large as the power savings achieved when DVFS is managed by the operating system. The second technique is rank-aware memory allocation, the goal of which is to power memory that the database system needs and avoid powering memory it does not need. We present experiments that show rank-aware allocation allows unneeded memory to move to low-power states, reducing memory power consumption.