Poster: Embracing NVM: Optimizing $B^𝜖$-Tree Structures and Data Compression in Storage Engines

Non-volatile memory (NVM) introduces a new class within the traditional storage hierarchy, combining attributes of both primary and secondary storage. These technologies offer latency approaching that of DRAM, though slightly higher\footnote{NVM technologies such as MRAM~\cite{slaughter2002fundamentals}, PCM~\cite{fong2017phase}, and RRAM~\cite{chen2020reram} each have distinct latencies.}, and significantly lower than traditional block storage devices. NVM is byte-addressable and provides persistence. Integrating NVM into various database and file systems has been the focus of extensive research. Efforts have centered on optimizing data structures like B-trees and LSM-trees, as well as enhancing components such as logging and recovery. However, a unified storage engine specifically designed to study the characteristics of the modern storage landscape, effectively utilizing and managing the potential of diverse memory and storage devices, does not currently exist. This work involves extending Haura, a general-purpose storage engine, to explore the benefits of the modern storage landscape. By integrating NVM into its storage stack, we aim to optimize data structures to leverage NVM's potential and employ compression techniques to minimize the memory footprint. This research is still in progress. However, the initial experiments integrating NVM into Haura demonstrated its strengths under sequential workloads and sensitivity to block sizes. Further modifications enable zero-copy deserialization and granular key-value access, fully leveraging NVM's byte-addressable characteristics.