Optimizing both performance and tail latency for B+tree on persistent memory

Abstract

B${}^{+}$-trees are widely used in databases and they have been optimized for persistent memory (PM) in recent studies. However, existing PM-oriented B${}^{+}$-tree designs are facing write performance penalties, high tail latency, and scalability issues, which are caused by three critical design limitations and the issues can be amplified on PM due to asymmetric write and read performance of PM: (1) node splits can lead to massive data migration; (2) frequent node splits can lead to high overhead of cascading modification; (3) node revision can lead to inefficient parallelism. In this paper, we propose a novel B${}^{+}$-tree-based index for PM with High write performance and Low tail latency, called HLTree, to solve the aforementioned issues and optimize both performance and tail latency for PM-oriented B${}^{+}$tree. First, HLTree employs a new node pre-split strategy to reduce the write overhead of legacy B${}^{+}$-tree designs. Second, HLTree decouples the structural modification operations from the critical path of the B${}^{+}$-tree and completes it asynchronously to reduce the overhead of cascading modification. Finally, HLTree optimizes optimistic version locks to reduce conflicts among readers and writers for lower latency and better scalability. Based on the evaluations conducted on Intel Optane DCPMM, compared with $\mu$Tree/SSB-Tree/Fast&Fair/FPTree, HLTree provides 1.06$\times$/2.38$\times$/2.16$\times$/1.55$\times$ read throughput and 1.50$\times$/2.28$\times$/2.13$\times$/1.58$\times$ write throughput on average, respectively. Moreover, HLTree reduces up to one order of magnitude lower of the 99.9th percentile tail latency.