Scaling Beyond DRAM with PMem without Compromising Performance (KE401/22-2)

PI: Alfons Kemper, Jana Giceva, Thomas Neumann (TU Munich)
Project Collaborator: Alexander van Renen, Lukas Vogel
Project website: To be announced

As DRAM reaches its scalability limits, it becomes necessary to rethink database systems for new storage technologies. The most promising is persistent memory, which was released in mid 2019. It has latencies and throughput almost within the same order of magnitude as DRAM, while being persistent and scaling to the sizes of SSDs at lower costs than DRAM. In this project we investigate possibilities for the incorporation of persistent memory into our state-of-the-art research database system, Umbra. As the successor of HyPer, it is perfectly suited for this project as a realistic (and invaluable) test bed for an end-to-end investigation to achieve and prove practical relevance of past foundational research. In particular, we are planning to investigate storage structures (LSM-Trees), index structures (B-Trees), logging and recovery as well as several performance optimisation techniques for persistent memory.