A Prolegomenon on OLTP Database Systems for Non-Volatile Memory

The design of a database management system’s (DBMS) architecture is predicated on the target storage hierarchy. Traditional diskoriented systems use a two-level hierarchy, with fast volatile memory used for caching, and slower, durable device used for primary storage. As such, these systems use a buffer pool and complex concurrency control schemes to mask disk latencies. Compare this to main memory DBMSs that assume all data can reside in DRAM, and thus do not need these components. But emerging non-volatile memory (NVM) technologies require us to rethink this dichotomy. Such memory devices are slightly slower than DRAM, but all writes are persistent, even after power loss. We explore two possible use cases of NVM for on-line transaction processing (OLTP) DBMSs. The first is where NVM completely replaces DRAM and the other is where NVM and DRAM coexist in the system. For each case, we compare the performance of a disk-oriented DBMS with a memory-oriented DBMS using two OLTP benchmarks. We also evaluate the performance of different recovery algorithms on these NVM devices. Our evaluation shows that in both storage hierarchies, memory-oriented systems are able to outperform their disk-oriented counterparts. However, as skew decreases the performance of the two architectures converge, showing that neither architecture is ideally suited for an NVM-based storage hierarchy.