Non-Volatile Memory (NVM) has emerged as an alternative to the next-generation main memories in recent years. NVM has the advantages of non-volatility, byte addressability, and high density. However, it also has some limitations compared to DRAM. For example, the read/write latency of NVM is higher than that of DRAM. The write endurance of NVM is also limited. That means replacing DRAM with NVM to construct an NVM-only system is not realistic. A more feasible way is to build an NVM/DRAM- based hybrid memory system. The B+-tree is a high-performance index structure that has been widely used in modern DBMSs. So far, most indices for hybrid memory are based on the B+-tree, such as FPTree, DPTree, LBTree, NBTree, and so on. They have proposed some effective ideas to improve the performance of the index, e.g., (1) just putting leaf nodes on NVM while maintaining all inner nodes on DRAM and (2) adopting unsorted leaf nodes. However, there remains potential for further improving the query performance of these indices.
To improve the performance of B+tree on hybrid memory, a research team led by Peiquan Jin published their new research on 15 October 2024 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
The team proposed a new index structure called HATree. The novel idea of HATree is to utilize the unused space of inner nodes to cache hot keys, thus improving query performance without using extra memory resources. To further improve the update performance of HATree, the team proposed to utilize the eADR technology supported by the third-generation Intel Xeon Scalable Processors to enhance HATree with instant log persistence, which results in the new HATree-Log structure. The experimental results show that the proposed index structure achieves higher search and update performance than previous works.
Future work can focus on modifying the read cache to be a read/write cache, implementing concurrent versions of the index.
DOI: 10.1007/s11704-023-3344-x