Burger-tree: A Three-Layer Cache-Conscious Tree Index for Persistent Memory

Persistent memory (PMEM) has become a new alternative of future main memories. Building efficient index structures for persistent memory has received much attention in recent years. We note that different nodes in a tree index have various read/write tendencies. Motivated by such an observation, we propose a new three-layer tree index called Burger-tree, which includes a top layer, a middle layer, and a bottom layer. The top layer is read-friendly and read-only, which is to accelerate search operations on the index. The middle layer is read/write friendly, while the bottom layer is write-friendly with an unsorted node structure. We present efficient node structures for the three layers, together with a new slot-forwarding algorithm to improve the efficiency of Burger-tree. With the three-layer structure, we demonstrate that Burger-tree can deliver high performance for read and write operations. We conduct experiments on real Intel Optane Persistent Memory and compare Burger-tree with WB-tree, FAST&FAIR, and TLBtree on YCSB workloads. The results suggest the efficiency of Burger-tree.