A highly write-optimized concurrent B[formula omitted]-tree for persistent memory

Abstract

Writes to persistent memory are considered expensive for the limited naive write performance of the device and the implementation of persist primitives. Write-optimized strategies are widely adopted in designing B+-Trees for persistent memory. In this paper, we concluded two major issues, including successive writes and recovery efficiency, on existing write-optimizing strategies. A highly write-optimized B+-Tree, HW-tree, is proposed to address these issues. HW-tree is constructed with three novel designs including distributed validation, log-free split, and concurrent recovery. Experimental results show that HW-tree can outperform its counterparts by large margins in almost all test workloads. The recovery time for reconstructing the tree can also be reduced by 88%.