WOPE: A write-optimized and parallel-efficient B[formula omitted]-tree for persistent memory

Abstract

Emerging Persistent Memory (PM) usually has the serious drawback of expensive write activities. Thus, existing PM-oriented B+-trees mainly concentrate on alleviating the write overhead (i.e., reducing PM writes and flush instructions). Unfortunately, due to the improper data organization in the sorted leaf node, existing solutions cause massive data migration when data inserting or node splitting occurs. In this paper, we propose a write-optimized PM-oriented B+-tree with aligned flush and selective migration, called WPB+-tree, to solve the above problems. WPB+-tree first adopts a buffer-assisted mechanism that temporarily stores the newly inserted data to reduce the overhead of entry shifts. Second, WPB+-tree employs a selective migration of node entries scheme to achieve less than half of the data migration when a node is split. Moreover, existing PM-oriented B+-trees usually employ a coarse-grained lock to avoid thread conflicts, which can severely degrade the concurrency efficiency. Thus, we further propose a fine-grained lock technique for WPB+-tree, namely, parallel-efficient WPB+-tree (WOPE), to improve the concurrency efficiency. We implement the proposed WPB+-tree and WOPE on Linux and conduct extensive evaluations with actual persistent memory, where WOPE achieves 23.5%, 30.7%, and 15.3% of performance improvement (insert, read, and scan) over the straightforward solutions (i.e., SSB-Tree, Fast&Fair, and wB+tree), and 10.1% of performance improvement over WPB+-tree, on average.