Time-series storage for MLC nonvolatile memories toward performance and reliability co-optimization

Recently, the wide deployment of surveillance systems such as driving recorders and sensor networks has highlighted the significance of time-series data, that are, the data indexed by timestamps. This trend has thus driven a rapidly increasing demand of high-performance, cost-effective storage and management systems for time-series data. In a typical use case, the storage system is accessed by range queries about the data created in a given time interval. Although there have been excellent work on the design and implementation of time-series databases, it remains a missing part on how to efficiently store and manage time-series data on a lightweight database based on energy-efficient nonvolatile memories (NVMs), such as flash memory and phase-change memory (PCM). This paper proposes Multi-level-cell Lightweight Time-Series Database (MLTSDB), which utilizes existing delta encoding technique and organizes the time-series data along with their associated differences (i.e., the 'delta') in the same set of NVM cells. The goal of such a design is to limit the amount of affected versions due to potential bit errors and to simplify the space management of NVMs on lightweight platforms like embedded systems.