Abstract: The persistent memory (PM) fascinates many researchers by its high scalability, byte-addressability and low static energy consumption which can contribute to the fusion of main memory and secondary memory. However，the interacting granularity between the volatile last level cache (LLC) and the non-volatile PM is normally 64B which is much larger than 8B, the granularity of the atomic updating operations provided by the main memory. Once an outage takes place during the process of transporting data from LLC to PM, data integration on the PM may be destroyed, which will lead to a breaking down of failure atomicity. Most researches are used to solve this problem by forcing the data persisted following some order implemented with flush and fence instructions which are always accompanied with large overhead. This overhead will be amplified especially in index updating which often leads to some transformation of index structures. In this paper, we focus on reducing the persisting overhead for ensuring the failure atomicity of updating a PM-based B\++-Tree. We propose a one direction shifting (ODS) algorithm based on the real data layout in the tree node as well as the node utility, the persisting overhead of different updating mode and the relation between different operations. This algorithm implements the in-place deletion to reduce the deleting overhead and takes advantage of the pits generated by in-place deletion to further reduce the shifting overhead of insertion. We redesign the rebalancing process of our persistent B\++-Tree according to the ODS algorithm. Experiments show that our proposed ODS tree can outperform the state-of-the-art PM trees on single and synthetic workloads.

Key words: persistent memory, index structure, failure atomicity, index updating, last level cache, persistent instruction