To content
Department of Computer Science

OCTO+: Optimized Checkpointing of B+ Trees for Non-Volatile Main Memory Wear-Leveling

Title

OCTO+: Optimized Checkpointing of B+ Trees for Non-Volatile Main Memory Wear-Leveling

Authors

Christian Hakert, Roland Kühn, Kuan-Hsun Chen, Jian-Jia Chen and Jens Teubner

Published

The 10th IEEE Non-Volatile Memory Systems and Applications Symposium (NVMSA)

Abstract

Steady deployment of byte addressable non-volatile memories (NVMs) as main memory in server class computers yields challenges on software. In order to overcome shortcomings, as for instance low cell endurance and high access latencies, working data can be kept in DRAM and continuously be checkpointed to the NVM. Although this reduces the impact of the NVM on usual execution, it shifts the endurance and latency issue to the checkpointing. Alongside widely studied generic wear-leveling solutions, we propose an application cooperative wear-leveling scheme for B+trees, that realizes an interplay of the application and the wear-leveling. We collect memory usage statistics during tree operations and dynamically choose a memory mapping between the DRAM footprint and the NVM checkpoint of the B+tree. In an experimental evaluation, we achieve 3×improvement in terms of memory lifetime.

Publication Log

August 2021

May 2021

Submitted to NVMSA '21