Data Management on Non-Volatile Memory: A Perspective

摘要

The large performance gap between main memory and secondary storage accounts for many design decisions of traditional database systems. With the upcoming availability of Non-Volatile Memory (NVM), which has latencies in the same order of magnitude as DRAM, is byte-addressable and persistent, a completely new type of technology is added to the memory stack. This changes some basic assumptions such as slow storage, block granular access, and that sequential accesses are much faster than random accesses. New ideas are therefore needed to efficiently leverage NVM. Although several new approaches can be found in the literature, the exact role of NVM is not yet clear. In this paper, we survey recent work in this area and classify the existing approaches. We focus on two key challenges: (1) integration of NVM into the memory hierarchy and (2) the design of NVM-aware data structures. We contrast the different approaches, highlight their advantages and limitations, and make recommendations.