Memory can consume a substantial amount of power in database servers, yet memory power has received considerably less attention than CPU power. Memory power consumption is also highly non-proportional. Thus, memory power becomes even more significant in the common case in which a database server is either not completely busy or not completely full. In this paper, we study the application of two memory power optimization techniques - rank-aware allocation and rate-based layout - to database systems. By concentrating memory load, rather than spreading it out evenly, these techniques create and exploit memory idleness to achieve power savings. We have implemented these techniques in a prototype database system called DimmStore. DimmStore is part of a memory power testbed which includes customized hardware with direct power measurement capabilities, allowing us to measure the techniques' effectiveness. We use the testbed to empirically characterize the power saving opportunities provided by these techniques, as well as their performance impact, under YCSB and TPC-C workloads. Under simple YCSB workloads, power savings ranged up to 50%, depending on load and space utilization, with little performance impact. Savings were smaller, but still significant, for TPC-C, which has more complex data locality characteristics.

The growing appetite for in-memory computing is increasing memory's share of total server power consumption. However, memory power consumption in database management systems is not well understood. This paper presents an empirical characterization of memory power consumption in database systems, for both analytical and transactional workloads. Our results indicate that memory power optimization will be effective only if it can reduce back-ground power through more aggressive use of low power memory idle states.