Research And Implementation Of NUMA-friendly Hierarchical Index Tree Based On Non-volatile Memory

Non-Uniform Memory Access(NUMA)architecture can break through the performance bottleneck of the memory bus.However,the multi-CPU NUMA architecture inevitably generates remote access under dynamic load.Frequent remote access operations with high latency will affect system performance.Non-Volatile Memory(NVM)has access performance close to that of Dynamic Random Access Memory(DRAM)and can persist data.However,its much higher access latency than DRAM will worsen the concurrent performance loss of NUMA remote access.Existing NVM-based key-value indexing techniques do not sufficiently address the performance degradation caused by frequent NUMA remote access.To solve the remote access performance problem of NUMA architecture in a high concurrency environment,a NUMA-friendly hierarchical index Tree(NFTree)was proposed.NFTree can effectively reduce NUMA remote access overhead by caching and migrating hotspot data.NFTree is divided into an index layer and a data layer.The index layer maintains an index B+tree in DRAM that uses high-speed DRAM to speed up keyword lookups.The data layer maintains several subtrees and logs in NVM for persistent storage of data.Each subtree corresponds to a leaf node of the index tree.NFTree reduces local NVM access by caching hotspot subtrees in DRAM and reduces remote NVM access by migrating most remote hotspot subtrees to the local NVM of the current CPU in the background.In addition,NFTree uses key-value separation technology to improve index storage efficiency and uses logging and copy-on-write technologies to reduce consistency overhead.Eventually,NFTree converts most remote NVM access to local NVM access and DRAM access,effectively reducing the remote access overhead of NUMA architecture.The NFTree system prototype is implemented on a real NUMA server.The read and write performance of NFTree was tested and compared with the B+tree indexes FAST&FAIR and Nap based on NVM.The test results show that the access performance of NFTree improves by 53%～97% compared with FAST&FAIR and 32%～72% compared with Nap under loads of reading,writing,and range querying.NFTree performs better when dealing with hot data.