Optimizing Main Memory Utilization of Columnar In-Memory Databases Using Data Eviction

Despite falling prices for main memory and increasing sizes, main memory is still a scarce resource in database systems. Optimizing main memory utilization is a major objective for main memory databases as more free memory can be used to improve performance or to store larger systems in the database. Several publications proposed separating frequently and less frequently accessed data (i.e., hot and cold data), handling both with different priorities or evicting cold data to secondary storage. However, most of these approaches are optimized for OLTP workloads. In contrast, this PhD project researches how to improve DRAM utilization for mixed workloads including both OLTP and OLAP queries by evicting cold data. As a first step, real-world database workloads are analyzed in order to determine characteristics of hot and cold data as well as aging effects. Also two possible approaches exploiting the results of the workload analyses are outlined.