CANRT: A Client-Active NVM-Based Radix Tree for Fast Remote Access.

This paper presents the first study of building a remote-accessible persistent radix tree, named CANRT. Unlike prior works that only focus on designing single-node tree structure for non-volatile memory, we focus on optimizing remote access performance for a persistent radix tree while minimizing the persistence overhead. Simply adopting server-reply paradigm will incur heavy server CPU consumption and hence lead to high operation latency under concurrent workloads. Therefore, we design a low-latency node-oriented read mechanism and a fine-grained lock-based write mechanism to minimize the server CPU involvement in the critical path. We also devise a non-blocking resizing scheme in CANRT. The extensive experimental results on commercial Intel Optane DC Persistent Memory platform show that CANRT outperforms the state-of-art server-centric persistent radix trees by 1.19x–1.22x and 1.67x–1.72x in read and write latency, respectively. CANRT also gains improvement of 7.44x–11.15x in terms of concurrent throughput under YCSB workloads.