Interference-aware Classification and Deployment of Database Workloads on Heterogeneous Memory Systems

Abstract

Over the recent years, the rapid growth of applications has led to a corresponding increase in the computational requirements and the volume of data. At the same time, traditional memory technologies based on DRAM technologies have reached their physical limits in terms of scalability and sustainability. As a result, a novel memory technology architecture has been introduced, based on heterogeneous memory systems, that integrate DRAM and NVM technologies. However, some several challenges arise regarding application placement and their management. In this work, it is conducted an extensive profiling analysis and characterization of database applications. The experimental analysis is performed both in an isolated computing system environment and in a interference-aware system using specific benchmarks that stress the resources of the system. Additionally, machine learning models are trained based on the data collected during the profiling analysis, and applications are classified into memory-bound, cpu-bound or bandwidth-bound depending on their performance. The models achieve prediction accuracy over 90%. Finally, conclusions can be drawn regarding the placement of applications in the heterogeneous system.