Rio: Fast B+-Tree Based on Remote Accessible Non-Volatile Memory

Abstract

Emerging Non-Volatile Memory (NVM) technologies, such as 3DXPoint, incorporating the merits of non-access speeds comparable to DRAM and storage-like persistence and capacity, hit the market in the near future. It is challenging to using NVMs in main memory way in view of NVMs’ durability, limited write endurance, and so on; as one of the crucial features of data structures designed for NVMs, data consistency needs to be taken care of explicitly so that the stored data in NVM can survive the system failure. However, data consistency comes at the cost of data concurrency, even in the distributed environment, especially for the ultra-low latency Remote Direct Memory Access (RDMA) technology. In this paper, we implement one variant Non-Volatile B+-Tree, Rio, based-on remote direct accessible NVM, preliminary experimental result shows that insertion throughput of Rio is circa 56% better than that of server-centric Non-Volatile B+-Tree.